In progress – view active session
Conference 13405
Physics of Medical Imaging
16 - 20 February 2025 | Town & Country B
Conference Co-Sponsor
Show conference sponsors +
View Session 
Information
Presenting student authors for this conference may be eligible for the Robert F. Wagner All-Conference Best Student Paper Award and the Physics of Medical Imaging Student Paper Award. View the award and application pages for details.
SPIE Medical Imaging Awards and Plenary
16 February 2025 • 5:30 PM - 6:30 PM PST | Town & Country B/C
View Full Details: spie.org/medical-imaging-awards-plenary
5:30 PM - 5:40 PM:
Symposium Chair Welcome and Best Student Paper Award announcement
First-place winner and runner-up of the Robert F. Wagner All-Conference Best Student Paper Award
Sponsored by:
MIPS and SPIE
5:40 PM - 5:45 PM:
New SPIE Fellow acknowledgments
Each year, SPIE promotes Members as new Fellows of the Society. Join us as we recognize colleagues of the medical imaging community who have been selected.
SPIE Harrison H. Barrett Award in Medical Imaging
Presented in recognition of outstanding accomplishments in medical imaging
13408-500
Development and translation of 3D ultrasound-based imaging systems for diagnostic and image-guided interventions
(Plenary Presentation)
16 February 2025 • 5:50 PM - 6:30 PM PST | Town & Country B/C
Show Abstract +
Our research has been focused on developing 3D US scanning devices that overcome the limitations of conventional US imaging methods. We have been developing and fabricating various mechanical external motorized fixtures that move a conventional US probe in specific patterns and used them in systems for image-guided prostate biopsy prostate, prostate and gynecologic brachytherapy, and focal liver tumour ablation. As well, we developed 3D US-based system for point of care diagnostic application such as whole breast imaging, carotid plaque quantification, and hand and knee osteoarthritis.
Our approach allows scanning the desired anatomy in a consistent manner, imaging a large volume, integration of any manufacturer's 2D US probe into our fixtures, and integration of machine learning methods for rapid diagnosis and guidance. This approach provides a means of using US images with any US system with a small additional cost and minimal environmental constraints.
All-Symposium Welcome Reception
16 February 2025 • 6:30 PM - 8:00 PM PST | Martini Lawn
View Full Details: spie.org/mi/welcome-reception
Join your colleagues on the lawn for food and drinks as we welcome each other to SPIE Medical Imaging 2025.
Monday Morning Keynotes
17 February 2025 • 8:20 AM - 10:30 AM PST | Town & Country A
View Full Details: spie.org/monday-morning-keynotes
8:20 AM - 8:25 AM:
Welcome and introduction
8:25 AM - 8:30 AM:
Award announcements
- Robert F. Wagner Award finalists for conferences 13405, 13409, and 13411
- Physics of Medical Imaging Student Paper Award
13405-501
To be determined
(Keynote Presentation)
17 February 2025 • 8:30 AM - 9:10 AM PST | Town & Country A
13409-502
Designing AI for the clinic: Internalizing the clinical task when developing deep-learning algorithms for medical imaging
(Keynote Presentation)
17 February 2025 • 9:10 AM - 9:50 AM PST | Town & Country A
Show Abstract +
Deep learning algorithms for image reconstruction, processing and segmentation are being developed and showing strong promise for multiple medical-imaging applications. However, medical images are acquired for clinical tasks, such as defect detection and feature quantification, and these algorithms are typically developed and evaluated agnostic to this clinical task. This talk will first discuss why clinical-task-agnostic evaluation of AI algorithms can be misleading, emphasizing the need for clinical-task-based evaluation of these algorithms. This discussion will lead to a new frontier in designing deep-learning algorithms that explicitly account for the clinical task of interest. We will see through examples how the clinical task can be incorporated within the design of these algorithms and how this then poises the algorithm for success in clinical applications. Throughout the talk, we will note how the rich field of model observers provides a mechanism to both design and evaluate AI algorithms for clinical tasks.
13411-503
To be determined
(Keynote Presentation)
17 February 2025 • 9:50 AM - 10:30 AM PST | Town & Country A
Break
Coffee Break 10:30 AM - 11:00 AM
Session 1:
Novel X-ray Sources and Systems
17 February 2025 • 11:00 AM - 12:40 PM PST | Town & Country B
13405-1
17 February 2025 • 11:00 AM - 11:20 AM PST | Town & Country B
Show Abstract +
The permitted input power density of rotating anode X-ray sources restricts the spatial X-ray image resolution. However, computed tomography would benefit from much smaller focal spots with equal output. Our group is proposing new tungsten microparticle targets that promise an order of magnitude improvement of the focal spot input power density, thus tripling the MTF limit.
Before investing, the limitations of classical technology should be known. Therefore, we modeled target erosion of rotating anodes that allows us to compute a new criticality parameter. In this context we also suggest a new correction factor for calculations of the patient X-ray dose. In conclusion, the specifications of X-ray tubes are justified. Unfortunately, the gain with increasing tube voltage is smaller than predicted by some volume heating models. Tungsten/rhenium coated carbon fiber reinforced anodes promise a few dozen percentage points MTF improvement, but cannot compete with newly proposed non-eroding microparticle-based targets.
13405-2
17 February 2025 • 11:20 AM - 11:40 AM PST | Town & Country B
Show Abstract +
Multi-source static Computed Tomography (CT) systems introduce novel opportunities for adaptive imaging. This work presents a fluence field modulation method using spotlight collimators, which block positive or negative fan angles of even and odd indexed sources, respectively. Spotlight collimators enable volume of interest (VOI) imaging by increasing relative exposure for those views which overlap with the VOI. To achieve high-quality reconstructions from sparse-view low-dose data, we introduce a generative reconstruction algorithm called Langevin Posterior Sampling (LPS), using a score-based diffusion prior and physics-based likelihood model to sample a posterior random walk. Simulation-based head CT imaging for stroke detection shows that spotlight collimators effectively reduce standard deviation and worst-case hallucinations in reconstructed images. Compared to uniform fluence, our approach significantly reduces posterior standard deviation, highlighting the potential for spotlight collimators and generative reconstructions to improve image quality and diagnostic accuracy of multi-source static CT, especially when using generative reconstruction algorithms.
13405-3
17 February 2025 • 11:40 AM - 12:00 PM PST | Town & Country B
Show Abstract +
This study evaluates the feasibility of reducing scatter and cone beam artifacts in dual-energy CBCT (DE-CBCT) by using an axially collimated carbon nanotube (CNT) x-ray source array with interlaced spectral filters. A benchtop CBCT scanner with an energy integrating flat panel detector (FPD), a CNT x-ray source array with 8 focal spots, and a rotating gantry was used. X-ray photons from 4 even-number focal spots were filtered by a low-energy (LE) filter, while the 4 odd-number spots by a high-energy (HE) filter. An external collimator was designed to confine the photons from each focal spot to illuminate only a section of the object. DE-CBCT images were collected by cycling the 8 beams sequentially during a full gantry rotation. An existing one-step materials decomposition and reconstruction method was modified for this specific imaging configuration. Phantom imaging studies showed significant reduced Cupping and cone beam artifacts, increased accuracy in materials quantification and enhanced image contrast, compared to a regular DE-CBCT.
13405-4
17 February 2025 • 12:00 PM - 12:20 PM PST | Town & Country B
Show Abstract +
Despite strict radiation exposure guidelines for pediatric patients due to their increased radiosensitivity, computed tomography (CT) remains essential in pediatric imaging because of its quick scan times and high spatial resolution. Spectral CT offers improved diagnostic quality by generating virtual monoenergetic images (VMIs) that enhance contrast and reduce artifacts without increasing radiation doses. However, most CT systems are optimized for adults, leading to suboptimal outcomes for children. To address this, a novel "double bowtie" filter combining traditional Teflon with K-edge material is proposed. This design optimizes spectral separability, enhancing signal-to-noise ratio (SNR) in pediatric imaging without increasing radiation exposure. Simulations show that this approach can significantly improve spectral CT performance for children, potentially reducing noise or radiation dose while maintaining image quality. Future studies will focus on clinical validation of this promising technique for pediatric diagnostics.
13405-5
17 February 2025 • 12:20 PM - 12:40 PM PST | Town & Country B
Show Abstract +
Breast mammography suffers from poor contrast between cancers and dense breast tissue. Likewise, while
breast microcalcifications play a crucial role in the early detection of breast cancer, mammographic techniques
struggle to differentiate between benign and potentially malignant calcifications non-invasively. We will present our
novel translatable method (not requiring high-resolution detector/mask pairs or spectral detectors) for a single shot single-mask approach. Such a mammographic phase contrast breast imaging can yield differential phase imaging to help distinguish low contrast masses while also yielding dark field information to help classify breast calcifications. Experiments with calcium oxalate and hydroxyapatite samples revealed distinct energy-dependent scattering profiles, particularly above 25 keV. This small angle scattering signals along with the differential phase contrast provides a promising avenue for non-invasive microcalcifiation classication as well as potential early detection of breast cancer in a single-shot mammographic setup. We will show non-spectroscopic and spectroscopic methods and compare classification benefits of each using breast tissue.
Break
Lunch Break 12:40 PM - 1:40 PM
Session 2:
Cone Beam CT
17 February 2025 • 1:40 PM - 3:00 PM PST | Town & Country B
13405-6
17 February 2025 • 1:40 PM - 2:00 PM PST | Town & Country B
Show Abstract +
The insertion of transpedicular screws is crucial in spine surgery but poses risks, as perforating the cortical bone can damage nerves and blood vessels. Integrating a C-arm cone beam CT system provides surgeons with visual feedback, guiding them during operations. However, physical effects that occur due to the interaction between the radiation emitted from the x-ray source and metal within the body, such as photon starvation, beam hardening and an increase in scatter radiation, can lead to artifacts that obscure images and reduce clinical value. Prior work optimized the tilt angle to reduce artifacts, improving image quality but not reliably predicting its sufficiency. This work aims to predict image quality based solely on scout views and a given trajectory. A phantom mimicking the cortical bone of the pedicles was created, alongside metrics to quantify metal artifacts. Several candidate predictors were developed and evaluated for their linear correlation with these metrics, achieving a Pearson correlation coefficient of up to 0.905.
13405-7
17 February 2025 • 2:00 PM - 2:20 PM PST | Town & Country B
Show Abstract +
In silico clinical trials and physical phantom studies are valuable for assessing new breast imaging technologies. Traditionally, ISTs and phantom studies use a detection task, to assess task-based performance from an observer, whether human or model. However, real clinical studies are more complex. Clinical breast imaging studies often require observers to both detect the presence of a suspicious lesion and rate the likelihood of its malignancy. Designing ISTs and physical phantom studies that accurately reflect clinical studies is challenging due to the difficulty in modeling malignant and benign lesions. One key feature for differentiating malignant from benign lesions is the lesion border, with malignant lesions often displaying spiculations. To better model clinical breast imaging studies, we explore the use of a classification task rather than a detection task. The task-based performance is evaluated using a CNN model, trained to distinguish between spiculated and non-spiculated masses.Our findings reveal significant variations in the conclusions drawn from ISTs with model observers depending on whether the task is lesion detection or classification.
13405-8
17 February 2025 • 2:20 PM - 2:40 PM PST | Town & Country B
Show Abstract +
In this work, we utilized the Cramér-Rao lower bound (CRLB) method to establish optimization framework for fast kV-switching (FKS) CBCT imaging with a spectral modulator, and quantitatively analyzed the impact of the different spectral modulators including materials, thickness and spatial frequencies on the FKS spectral imaging performance.
13405-9
17 February 2025 • 2:40 PM - 3:00 PM PST | Town & Country B
Show Abstract +
This study presents a novel deep learning approach for noise reduction in ultra-low dose (ULD) dental cone beam computed tomography (CBCT). The method pre-trains a U-Net model on phantom datasets and fine-tunes it with simulated datasets, balancing noise reduction and anatomical structure preservation. Qualitative evaluation shows that the proposed method maintains soft tissue details while reducing noise effectively. Quantitative analysis demonstrates consistent improvements in the peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) across phantom and simulated datasets compared to ULD images. This approach shows promise for clinical implementation of deep-learning-based noise reduction, potentially enabling diagnostic-quality CBCT scans at extremely low radiation exposure.
Break
Coffee Break 3:00 PM - 3:30 PM
Session 3:
Detectors
17 February 2025 • 3:30 PM - 5:30 PM PST | Town & Country B
13405-10
17 February 2025 • 3:30 PM - 3:50 PM PST | Town & Country B
Show Abstract +
This study introduces an innovative detector design – Coded Sensing (CS) SPECT, that integrates the geometry of both conventional pinhole collimator and coded aperture. This design not only improves sensitivity by two orders of magnitude compared to clinical SPECT, but also keeps directional information and maintains good spatial resolution.
We built a prototype CS detector module using 8x8 crystal array. Each crystal element presents an alternating pattern of GAGG scintillator and acrylic. The depth of interaction (DOI) information is obtained through dual-ended readout system. The detector performance was evaluated using different radioactive sources: Na-22, Tc-99m, Lu-177, and Pb-212. It demonstrated the energy resolution of the detector is ~10%, and it showed capability of working under a broad energy spectrum from 70 to 511 keV.
We derived the system response functions based on analytical ray-tracing model, and validated the results through Monte Carlo simulation. A preliminary imaging study was carried out using three Na-22 point sources. The image was successfully reconstructed with the derived system matrix, and the source at different locations can be clearly resolved.
13405-11
17 February 2025 • 3:50 PM - 4:10 PM PST | Town & Country B
Show Abstract +
In recent years we have seen an increased interest in differential phase-contrast imaging, where one extracts two additional signals from the scan: phase and dark-field. To implement the technique in a clinical setting, one needs to solve the challenge of the analyser grating used to resolve the interference pattern of the interferometer used in the technique.
This work aims to demonstrate that through charge-cloud fitting, sub-pixel resolution can be obtained resulting in a high-resolution CT-detector, with a potential resolution on the micrometer scale, high enough to directly resolve the interference pattern mentioned above.
To demonstrate this we perform measurements using a prototype with 14 µm pixel pitch and compare to Monte-Carlo simulations. We perform DAC-sweeps as well as a readout mode on the prototype that let us detect single photon interactions which, due to the charge-sharing between pixels, allows us to perform curve-fitting to estimate the interactions position.
Initial results are promising, with simulations and measurements agreeing very well, indicating a sub-pixel resolution.
13405-12
17 February 2025 • 4:10 PM - 4:30 PM PST | Town & Country B
Show Abstract +
In digital radiography imaging, stacked dual-layer flat-layer detectors enable single-exposure spectral imaging for a variety of applications, such as bone and tissue separations and super-resolution imaging. Dual-layer detectors can also improve the detective quantum efficiency (DQE). In this paper, we develop an aligned dual-layer detector (ADD). To maintain a uniform registration and magnification between images acquired from the two layers, we stack two layers so that the distance between the layers is as small as possible. We then conduct subpixel registration to estimate a subpixel-accuracy translation and then transform the images acquired from the lower layer based on the Fourier shift theorem and an affine transformation with a cubic interpolation. By using developed prototypes of ADD, we experimentally observe the DQE performance of the convex combination of the upper and lower images, and compare the IQFInv values. Finally, we discuss an issue on alleviating anti-scatter grid artifacts in ADD.
13405-13
17 February 2025 • 4:30 PM - 4:50 PM PST | Town & Country B
Show Abstract +
Amorphous Selenium (a-Se) has been extensively studied as a direct conversion detector material for x-ray
imaging. We present the evaluation of a hybrid a-Se/CMOS detector, fabricated by thermal evaporation of the
active layer directly onto the functional readout ASIC. In this study, the RD53C/ITkpixv2 front-end chip with
50x50 μm pixel pitch, developed for high-energy physics tracking detectors, is used. We demonstrate the detection
of both low-energy beta electrons and x-rays with the hybrid a-Se/RD53C detector, and study the impact of the
operation bias voltage across the a-Se layer on the hit counting rate and signal charge as represented through
the time-over-threshold.
13405-14
Optimizing parylene and photoconductor thickness in indirect conversion amorphous selenium detectors
17 February 2025 • 4:50 PM - 5:10 PM PST | Town & Country B
Show Abstract +
Amorphous selenium (a-Se) provides an opportunity for a low-cost, large-area avalanche photodetector for use in indirect conversion detectors. However, its bandgap of 2.2 eV reduces the response at long wavelengths, specifically the 550 nm green light emitted by CsI:Tl scintillators, limits its application. Incorporating tellurium into the a-Se conversion layer is known to reduce the bandgap and increase sensitivity at these longer wavelengths. This group has proposed utilizing a Se-Te conversion layer with a parylene hole blocking layer, with results demonstrating a need for optimization of layer thicknesses to achieve high sensitivity, reasonable leakage, and low lag and ghosting. In this study, we evaluate the effects of varying the parylene layer thickness and the photodetector conversion layer for single pixel a-Se devices, then implement it in a Se-Te device. conventions used for other insulating blocking layers, and a trade-off between low leakage and photoresponse for parylene layer thickness. We then demonstrate that by using a thin layer of parylene combined with a thin layer of Se-Te, high conversion efficiency at low voltages can be obtained.
13405-15
17 February 2025 • 5:10 PM - 5:30 PM PST | Town & Country B
Show Abstract +
Research on "Hybrid" active matrix flat panel imagers (AMFPIs) has shown their potential to combine direct and indirect imaging for digital radiography. These imagers use an amorphous selenium (a-Se) layer coupled with a CsI:Tl scintillator, potentially surpassing current detectors in detective quantum efficiency (DQE). The study investigates doping a-Se with tellurium (Te) to improve OQE to CsI:Tl emissions and evaluates a 6.5 x 6.5 cm² Hybrid AMFPI prototype. The Hybrid system demonstrated significant x-ray sensitivity increases (42% for RQA5 and 91% for RQA9) compared to direct AMFPI, with a tenfold OQE improvement. Although the Hybrid's modulation transfer function (MTF) was lower than that of the direct AMFPI, it outperformed the indirect component alone. The Te-doped Hybrid achieved the highest combined MTF and DQE for an energy-integrating imager under RQA9 conditions. Future work will focus on co-doping with arsenic to reduce lag for real-time imaging.
Posters - Monday
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
View Full Details: spie.org/mi/monday-poster-session
Conference attendees are invited to attend the SPIE Medical Imaging poster session on Monday evening. Come view the posters, enjoy light refreshments, ask questions, and network with colleagues in your field. Authors of poster papers will be present to answer questions concerning their papers. Attendees are required to wear their conference registration badges.
Poster Presenters:
Poster Setup and Pre-Session Viewing: 10:00 AM - 5:30 PM Monday
- In order to be considered for a poster award, it is recommended to have your poster set up by 1:00 PM Monday. Judging may begin after this time. Posters must remain on display until the end of the Monday evening poster session, but may be left hanging until 1:00 PM Tuesday. After 1:00 PM on Tuesday, posters will be removed and discarded.
https://spie.org/MI/Poster-Presentation-Guidelines
Poster groupings are listed below by topic.
Posters: CBCT
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
View poster session description and guidelines above.
13405-59
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This study proposed an energy-modulated scatter correction method for dual-layer flat-panel detectors (DL-FPD) based cone-beam CT (CBCT) imaging. By establishing a signal distribution model for the two detector layers, the X-ray scatter signals from both layers were analytically extracted. Validation through Monte Carlo simulations and phantom experiments showed that this approach significantly reduced scatter artifacts in both low-energy and high-energy CBCT images from DL-FPD. Quantitatively, it achieved a 77% reduction in non-uniformity for low-energy images and a 66% reduction for high-energy images, along with improved material decomposition accuracy.
13405-60
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Cardiovascular disease remains a leading cause of global mortality, driving innovation in cardiac imaging. Interior tomography significantly reduces radiation exposure, making it ideal for cardiac applications. A Triple-Source CT (TSCT) capable of performing interior tomography of the heart using three simultaneous imaging chains shows promise for rapid, low-dose cardiac imaging. However, conventional analytic reconstruction algorithms struggle with severe artifacts due to data truncation and sparse-view sampling. We propose a novel interpolation-based dual-task network for deep reconstruction of cardiac CT images in TSCT. Our approach employs two sub-networks that address sparse-view and truncation artifacts independently, leveraging distinct ground truths to constrain each sub-network during training. Additionally, we utilize two interpolation methods to complete the sinogram as a prior input. Experimental results from imaging a porcine heart with 84 views and a 50% truncation ratio show our method effectively suppresses artifacts, achieving 74.1% improvement in RMSE and 10.9% in SSIM compared to FBP.
13405-61
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Compared to single-energy computed tomography (CT), dual- or multiple-energy cone-beam CT (CBCT) has potential of offering better image quality and material differentiation capability. However, an accurate and fast scatter estimation is highly demanded, as the X-ray scattering influences imaging quality, resulting in inaccurate material decomposition and image artifacts. The linear Boltzmann transport equation (LBTE) is considered to be a fast and accuracy approach for scatter estimation. In this work, we introduce a new label dimension in LBTE (LBTE-L) and developed a unified and highly efficient scatter estimation method, which can calculate scatter signals at multiple different spectra in a single computation. We validate its effectiveness and accuracy by comparing it with the Monte Carlo Method and by applying scatter correction on the actual data measured in a spectral CBCT tabletop system.
13405-62
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Conventional cephalometric radiographs, such as lateral and posteroanterior (PA) cephalometric radiographs, are widely
used in orthodontics and oral maxillofacial surgery to evaluate bone and soft-tissue morphology. However, they often suffer from structural overlap and inadequate head positioning during the scan, which limits their diagnostic utility and affects treatment decisions. In this work, we present a new method for generating digital cephalometric radiographs from 3D images acquired with a dental CBCT system. With the proposed method, both digital lateral and PA cephalometric radiographs can be generated from 3D images obtained with a single CBCT scan. Our preliminary results reveal that the proposed method can yield cephalometric radiographs with improved contrast and landmarks of clinical relevance.
13405-63
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
The inherent limitations of cone beam CT (CBCT) compromise its performance for image-guided radiation therapy (IGRT), particularly in terms of accuracy, uniformity, and most importantly, low soft tissue contrast. A multisource CBCT (ms-CBCT) using a carbon nanotube (CNT) X-ray source array was developed to overcome these limitations while maintaining the essential advantages of CBCT. Brain imaging studies show that the CNT-based ms-CBCT significantly improves soft tissue contrast compared to conventional CBCT, allowing for easy differentiation of various soft tissues. The imaging quality is also comparable to that of multidetector CT (MDCT). Combining the advantages of both CBCT and MDCT, the CNT-based ms-CBCT could become the next-generation imaging tool for IGRT systems, potentially providing both patient alignment and tumor monitoring.
13405-64
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Metal artifacts in dental CBCT often obscure the tissues adjacent to metal implants, reducing the accuracy of CBCT examinations. This study proposes a novel dual-domain deep learning algorithm guided by multi-scale information to address metal artifacts in dental CBCT. Our approach combines physics-based imaging models with data-driven models, using wavelet transform to obtain different wavelet components as prior information for the network. The method comprises a multi-scale correction network, an image enhancement network, and a feature-map fusion network. By leveraging the unique feature-map and content of different frequency components, the network is guided to correct metal-contaminated sinograms. The corrected dual-domain feature-maps are then fused, promoting mutual learning between the sinogram domain and the image domain. Experimental results show that our method effectively removes metal artifacts. Compared to uncorrected images, the root mean square error (RMSE) is reduced by 73.7%, while the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) are improved by 45.2% and 18.6%, respectively.
Posters: Image-guided Intervention and Radiotherapy
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
View poster session description and guidelines above.
13405-65
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Pre- and post-treatment evaluation for treatment of spinal deformity with MAGEC rods is crucial, particularly for assessing respiratory volumes. However, presence of MAGEC rods precludes MRI scanning. Therefore, finding a method to measure regional respiratory volumes in such patients is essential for optimized clinical care. We introduce a novel method using pre-operative dMRI and pre- and post-operative CXR to predict post-operative lung volumes via a neural network. With a dataset of 49 pediatric patients with thoracic insufficiency syndrome, we demonstrated feasibility of predicting post-operative lung volumes with ~10% relative percentage error. Future studies with more datasets can achieve better results.
13405-66
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Improved dose monitoring and management using accurate, real-time deep neural network (DNN) prediction of scatter-dose distributions during fluoroscopically-guided interventional procedures will facilitate the minimization of staff risk. The effect of distribution training-set parameters such as its size, its spatial resolution, the dropout of data near the isocenter and logarithmic data compression on the accuracy of DNN prediction was investigated. The ground-truth dataset used for DNN training was generated via EGSnrc Monte-Carlo simulation and expanded by linear interpolation. The DNN input layer is a 1x5 vector whose elements are those factors which influence the shape of the procedure-room scatter distributions. With judicious selection of training set parameters, the DNN scatter-dose predictive model achieves a mean absolute percent error (MAPE) of approximately 5% when compared to ground truth. This DNN method can be incorporated into our Scatter Display System (SDS), which provides a real-time, color-coded visualization of procedure-room scatter distributions.
13405-67
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This study investigates the effects of metal artifact avoidance (MAA) through tilted trajectories and metal artifact reduction (FS-MAR) on intraoperative CBCT image quality in anterior cervical discectomy and fusion (ACDF) surgery. A high-fidelity phantom was used to assess how varying C-Arm tilt and applying MAR influence the visibility of the gap between the implant and the caudal vertebral surface. Gap visibility was measured using the contrast-to-noise ratio (CNR). Results showed that gap-bone CNR improved by a factor of 3.4 with MAA and by a factor of 2.5 with MAA+MAR. While MAA significantly enhanced CNR, MAR did not notably improve gap visibility due to secondary disturbances. Optimizing tilt angles can substantially improve image quality, though MAR, while reducing streaking artifacts, does not enhance gap visibility. Clinicians can use contextual knowledge to select tilt angles that prioritize critical areas, with future work needed to refine MAR techniques and validate these findings clinically.
13405-68
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Increased accessibility and recent developments in positron emission tomography (PET) detector technology have enhanced PET's role in radiotherapy treatment planning. This study investigates the efficacy of silicon photomultiplier (SiPM) PET systems for improving treatment volume delineation for radiotherapy. The study used a modified NEMA IEC Body Phantom filled with fluorodeoxyglucose (FDG) to simulate radioisotope uptakes, imaged using both a digital SiPM PET scanner and a non-digital PET scanner. Results show distinct differences in treatment volumes determined using the two systems, with the digital system demonstrating more comprehensive and accurate coverage of the treatment target volume.
13405-69
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Ischemic stroke is the leading cause of long-term disability in the United States. Fortunately, the landscape of stroke patient management has been changing by endovascular mechanical thrombectomy (EVT) in recent years. EVT is an interventional procedure to remove a stroke-causing thrombus (clot) from a cerebral artery to induce recanalization. It stands to reason that further improvements of EVT in safety and efficacy will continue to improve stroke outcomes. One of the key to improve EVT is to assess the brain perfusion in the interventional suite, before and after the EVT. For the purpose, we have been developing a novel method called IPEN for Intra-intervention PErfusion using a standard x-ray angiography system with No gantry rotation. The IPEN quantitatively assesses the brain perfusion on a region basis using a series of x-ray angiography images taken from one or two view angles. Simulation studies have confirmed that the IPEN works very well once the orientation of the head relative to the x-ray angiography system is known. This work aims just that.
13405-70
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
An FEM-based registration method has been developed for image fusion to facilitate treatment planning in C-arm CBCT guided histotripsy. This model is based on liver and gallbladder segmentations and is being evaluated in a preclinical in vivo porcine model. Deep learning models for human CT segmentation exist but are not generalizable to swine. In this study, a set of deep learning networks were created for the automatic segmentation of swine livers from CT, using different ratios of human and swine training data, and with and without randomizing the final layer of pretrained weights from a human segmentation model. Results show that automated segmentations from the highest performing network can be used to register images with accuracy comparable to registration from FEM models created from manual segmentations.
13405-71
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
In this study, we proposed a multi-scale attection U-Net-based generative adversarial network (MSA-CGAN) with a patch-content (PC) limit loss to accurately generate synthetic CT (sCT) images for adaptive radiation therapy (ART). The generator of the MSA-CGAN had multi-scale blocks and attention gates for securing the diversity in extracted features and delivering the pertinent information suitable for a synthetic task. The PC limit loss was calculated for the patches, which were identically selected for both of the source and fake images, and the patch selection condition was restricted to include an object in the patch. The results showed that the proposed model successfully synthesized CT images with high accuracy, and the performance of the MSA-CGAN with the PC limit loss was superior to that of the other techniques. In conclusion, the proposed model is able to provide accurate sCT images for the ART.
Posters: Image Reconstruction
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
View poster session description and guidelines above.
13405-72
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
In this work, we proposed the Projection Embedded Schrödinger Bridge (PESB) for CT sparse view reconstruction. PESB constructs Schrödinger Bridges between the distribution of Filtered Back-Projection (FBP) reconstructed images and the distribution of clean images conditioned on measured projections. By embedding projections into the marginal conditions, data consistency is inherently incorporated into the generative process. Experimental results validate the effectiveness of PESB, demonstrating its superior performance in CT sparse view reconstruction compared to several diffusion-based models.
13405-73
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Incorporating Time-of-Flight (TOF) information into Positron Emission Tomography (PET) imaging has been demonstrated to enhance the quality of reconstructed PET images and reduce noise. However, the realization of this enhancement is challenged by the substantial memory requirements of model-based deep learning reconstruction methods. To address this issue, we introduce a new model-based deep learning approach, LM-SPD-Net, designed for TOF-PET reconstruction from list-mode data in this study. Experimental outcomes reveal that the proposed method exhibits superior performance compared to current leading TOF-PET list-mode reconstruction algorithms, further reducing the spatial and temporal consumption required for reconstruction. This advancement paves a new avenue for the application of deep learning technology in TOF list-mode data, highlighting its potential to significantly improve PET imaging efficiency and accuracy. Our findings suggest that LM-SPD-Net offers a novel idea to the challenges faced in TOF-PET imaging.
13405-74
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Inverse rendering (IR) is an emerging field that facilitates the calculation and propagation of gradients of 3D objects through images. This study introduces a novel framework for high-resolution IR positron emission tomography (PET) reconstruction based on Dr. Just-in-time (Dr.JIT), a physically-based differentiable rendering compiler. The IR method allows for efficient auto-differentiation (AD) of voxel intensities with respect to each image parameter, enabling the incorporation of complex models, including complete scatter simulation and attenuation effects, within a unified framework that enhances overall accuracy and image quality. We simulated XCAT brain phantom data and reconstructed the coincidence data using objective functions with Least Squares Estimation (LSE), Weighted Least Squares Estimation (WLSE), and Maximum Likelihood Expectation Maximization (MLEM), both traditionally and AD-based. Results demonstrate that the AD-WLSE method offers the highest image resolution and quality, showcasing the platform's potential in PET image reconstruction.
13405-75
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Computed tomography (CT) imaging faces challenges in the intestine, due to peristalsis and content. Recently, a second-generation AI iterative reconstruction (AIIR2.0) algorithm was developed to address the shortcomings of its predecessor, AIIR1.0. This study briefly summarizes the advances of AIIR2.0, then, using colorectal cancer (CRC) as an example, aims to comprehensively evaluate the effectiveness of AIIR2.0 in abdominopelvic CT imaging. A dataset of CRC patients was established in this study, containing complete CT raw data from 17 patients. Compared to AIIR1.0, AIIR2.0 significantly reduces processing time (1110.3 s vs. 206.9 s), improving throughput and operational efficiency. AIIR2.0 offers a broader range of noise textures with more variable SNRs and CNRs across strength levels. Notably, AIIR2.0 provides more natural textures and details, addressing the “waxy” appearance at a high strength level. In conclusion, AIIR2.0 is more efficient for clinical diagnosis in abdominopelvic CT compared to AIIR1.0.
13405-76
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Decades of x-ray CT research focuses on iterative reconstruction and denoising methods to alleviate constraints on data sampling and ionizing radiation dose. Multi-channel CT imaging applications (multi-energy, dynamic) remain an active area of research because the relationships across channels enable highly effective data undersampling and reconstruction. Now, deep learning (DL) reconstruction methods are at the forefront of CT research. A key to the success of DL reconstruction methods is their ability to learn data-specific prior information as a supplement for mathematical optimization methods. Previously, we demonstrated how the split Bregman optimization method can be combined with supervised learning and simulated CT data to enable volumetric, projection and image domain (dual domain) reconstruction of real mouse micro-CT data. Here, we extend this framework in three key ways: (1) we apply it to multi-channel, time-resolved CT reconstruction (3D + time); (2) we replace the original convolutional neural networks with 3D vision transformers (ViTs) with several modifications; and (3) we integrate our DL model with our open-source reconstruction tools.
13405-77
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
In this work, we investigate accurate image reconstruction of multiple basis images with the basis-region data model and volume conservation constraint. Interest exists for reconstructing multiple (more than two) basis images directly from dual-energy CT (DECT) data, as scanned objects may contain K-edge materials and/or metal implants. We have previously proposed a basis-region data model for reconstructing multiple basis images directly from DECT data by partitioning the image array into basis regions and allowing no more than 2 basis materials in each basis region. Here, we extend the previous work by incorporating the volume conservation constraint and thus allowing up to 3 basis materials in each basis region and developed a modified non-convex primal-dual (NCPD) algorithm for reconstructing multiple basis images. Results from simulated- and real-data studies suggest that the modified NCPD algorithm can accurately recover multiple basis images from consistent noiseless data and stably reconstruct multiple basis images and virtual monochromatic images (VMIs) with little to no visual artifacts from real dual-energy data.
13405-78
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
In computed tomography (CT), transverse truncation occurs when the X-ray beam's transverse field-of-view (FOV) is smaller than the subject's size. While the FOV can be expanded by offsetting the detector, this adjustment may still be insufficient for large subjects or small detectors. Existing research primarily focuses on image reconstruction within the FOV. In contrast, our study aims to achieve accurate reconstructions both within the FOV and across a significantly larger region. We formulate the reconstruction problem from truncated CT data as a convex optimization problem, incorporating hybrid constraints on total variation (TV) and L1-norm within different image regions. We propose the TVL1 algorithm for accurate reconstructions within a region substantially larger than the FOV. Through simulations and real-data experiments involving various truncation levels using the offset detector setup, we demonstrate the algorithm's robustness and performance. Our findings reveal the TVL1 algorithm's effectiveness in reconstructing images within and beyond the FOV under different practical truncation scenarios.
13405-79
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
We propose new pixel-wise hyperparameter fine-tuning methods for penalized-likelihood tomographic reconstruction. Initially, all hyperparameters are set for the entire image, either manually or using an automated method such as a deep learning-based approach. To adjust the hyperparameter controlling the shape of the penalty function, we employ a method based on patch similarities. To adjust the smoothing parameter that balances the data-likelihood and regularization terms, we propose a new method that transforms the cumulative histogram obtained from the standard deviation of the estimated image at each iteration. Simulation results indicate that adjusting both the control and smoothing parameters using our proposed methods improves overall reconstruction accuracy, as measured by several image quality assessments.
13405-80
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Neural fields have recently emerged as a promising approach for solving imaging inverse problems. In the context of computed tomography (CT), neural field resembles an iterative reconstruction method but leverages a deep neural network optimized through gradient descent. In this work, we propose a novel method for regularizing neural fields by incorporating ray directions from arbitrary viewpoints. Our approach promotes smoothness along the ray samples from multi-views, enabling the network to learn more informative signals and represent accurate geometry. Experimental results on cone-beam CT dataset demonstrate the effectiveness of our method in improving sparse-view CT reconstructions, with notable reductions in streak artifacts and enhanced representation of anatomical features. In summary, this work presents a simple yet effective regularization to improve 3D tomographic imaging from sparse X-ray projections.
13405-81
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
For calibration-based magnetic particle imaging (MPI) reconstruction, frequency selection is an important step as it can significantly reduce the computational burden. To suppress the spectral leakage problem of current SNR methods, we proposed a targeted Focused-SNR method in order to select more useful frequency components for high-quality image reconstruction. With the proposed method, a balance between reducing the computational burden and guaranteeing the reconstructed image is achieved. The experiments are compared using public Open-MPI dataset, and the results show that the proposed method is able to reconstruct high-quality image and have higher robustness when threshold changes.
13405-82
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This study presents a novel methodology for optimizing CT protocols using Virtual Imaging Trials (VITs) and reinforcement learning. Computational phantoms with lung disease were imaged using a validated CT simulator and reconstructed with a novel CT reconstruction toolkit. The optimization parameter space included tube voltage, tube current, reconstruction kernel, slice thickness, and pixel size. A Proximal Policy Optimization (PPO) agent was trained to minimize the Root Mean Square Error (RMSE) between reconstructed images and ground truth phantoms. The reinforcement learning approach achieved a local minimum RMSE within 1 HU of the absolute minimum RMSE, with 80.5% fewer steps compared to an exhaustive search. This methodology demonstrates a robust and flexible framework for CT protocol optimization that is adaptable to various image quality metrics.
13405-83
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Limited-angle (LA) dual-energy (DE) cone-beam CT (CBCT) is considered an ideal solution to achieve fast and low-dose DE imaging on current CBCT scanners without hardware modification. However, its clinical implementations are hindered by the challenging image reconstruction from insufficient projections. In this work, we aim to perform optimization-based image reconstruction for LA-DECBCT with high computation efficiency. A structural similarity-based regularization term is introduced into the joint image reconstruction of DECBCT for suppression of LA-artifacts based on the facts that i) the DECBCT images share the same anatomical structures, and ii) the complete anatomical information is acquired in the LA-DECBCT projection data. The proposed iterative reconstruction method is evaluated using a Catphan phantom study and a digital phantom study, showing its feasibility for efficient image reconstruction in LA-DECBCT.
13405-84
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Two versions of fan-beam FBP reconstruction algorithms without spatially varying backprojection weight were proposed in Ref. [1] and [2], respectively. The two algorithms have strong resemblance and difference. In this work, we provide an approach to demonstrate how to reach algorithm-2 with the knowledge in Ref. [1] which demonstrates the resemblance between them. Both algorithms have channel and view contributions to the resulting image. However, in the full scan problem, the view contributions of algorithm-1 are zero while the view contributions algorithm-2 are non-zero. We comment also that algorithm-2 cannot be used in a reduced short scan protocol while algorithm-1 is applicable. We propose a hybrid convolution scheme to avoid differentiation along the channel direction for both algorithms which can improve the spatial resolution in the resulting image.
13405-85
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Sparse-view CT imaging reduces radiation dose and scan time but often results in poor image quality. Traditional reconstruction methods mainly use CT characteristics as priors, overlooking the potential of other modalities like MRI. Our study introduces a novel approach leveraging wavelet transform to extract MRI features as priors for sparse-view CT reconstruction via a dual-domain iterative network. By decomposing CT images into high and low-frequency components, we utilize MRI-derived priors to recover high-frequency details. This physics-based, data-driven model integrates information from both image and sinogram domains, enhancing interpretability. Experimental results demonstrate significant artifact reduction and improved image quality, showcasing the benefits of multi-modality information in sparse-view CT imaging.
Posters: Artificial Intelligence
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
View poster session description and guidelines above.
13405-86
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This study optimizes a deep learning-based U-Net++ model for segmenting internal carotid artery and vertebral artery in computed tomographic angiography (CTA) images, aiming to reduce patient radiation exposure by eliminating additional bone scans. Using dual energy and direct subtraction CTA data from 70 patients, we evaluated various pruning range of the U-Net++ model. Results show that optimal pruning range maintains segmentation performance while significantly reducing computational time. The findings suggest the U-Net++ model's potential to improve diagnostic accuracy in cerebrovascular disease by accurately segmenting cerebral vessels in CTA images.
13405-87
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
CT scout images help plan the scan range and optimize radiation dose, customizing exams to each patient. For these exams to be fully personalized at the organ level, the precise location of the patient's relevant anatomy is of utmost importance. We explore deep learning-based organ segmentation methods for frontal and lateral scout images using variations of a U-Net model with a pretrained ResNet-50 encoder backbone. We demonstrate that automatic organ segmentation is a feasible and attractive task with broad applications to CT exam planning, dose optimization, surgical planning and guiding procedures.
13405-88
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Chest radiography with dual-energy subtraction (DES) enhances the detectability of lung cancer and inflammatory diseases but has limitations like deteriorated image noise and increased radiation dose. Previously, we developed an artificial intelligence-based DES (AI-DES) system to address these issues, but it required labor-intensive processing and produced images with significant artifacts. In this study, the AI-DES system was enhanced by introducing a new data-loading method and implementing automatic processes in preprocessing and determining weighted factors. Weighted factors for soft-tissue images were optimized using histogram analyses, while those for bone-enhanced images were determined through linear approximation. The fully automated AI-DES system effectively highlighted pulmonary lesions in soft-tissue-enhanced images and revealed bone lesions and calcifications in bone-enhanced images. Loading image data as floating-point numbers instead of 8-bit integers significantly reduced artifacts and improved overall image quality. The system’s full automation and effective lesion enhancement strongly support its clinical applicability.
13405-89
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Recently, several deep learning (DL)-based methods have been suggested to reduce the noise in low-dose computed tomography (LDCT) images. However, those approaches often require sufficient pairs of LDCT and normal-dose CT (NDCT) images, which are difficult to acquire in clinical practice. On the other hand, most existing DL-based methods suffer from generalization issues because the network is trained with one or a few specific dose levels. To overcome the aforementioned problems, we propose a novel denoising framework that can be easily generalized to various dose levels while using only NDCT images. Furthermore, we mathematically derive a noise-variance calibration module for network training to handle the over-smoothing in network’s output. Once the network is trained, we can adaptively reduce the noise in test LDCT images of different doses. The proposed method shows superior performance in preserving noise textures while achieving the best scores in the quantitative assessment that evaluates perceptual quality.
13405-90
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This paper proposed a quality enhancement method for radiographic X-ray images, which providing ability of interpreting the pixel-level mapping functions. Extensive experiments are performed on collected X-ray clinical images, which indicate the proposed method has better performance than the state-of-the-art DL methods. The resulted images provide promising contrast enhancement of the X-ray images regional and globally.
13405-91
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
In dental CT images, metal artifacts hinder accurate analysis due to their high attenuation coefficient. Traditional MAR methods and deep learning techniques often suffer from residual artifacts and z-axis directional discontinuities due to domain gap between simulated and real data and processing in two dimensions. We propose a novel method using a self-data augmentation and z-continuity (ZC) module to address these issues. By generating training data based on clinical characteristic and using morphology layer to manage z-axis continuity, the method effectively reduces artifacts and improves image quality. Experimental results demonstrate its effectiveness in artifact reduction and anatomical preservation.
13405-92
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This paper presents a novel approach using a pre-trained Score-based Generative Model (SGM) for low-dose CT(LDCT) denoising. Our novel approach bypasses the initial stages of the SGM inference process by directly utilizing LDCT images as if they were generated in these early steps. To seamlessly integrate LDCT images into the inference process, the Langevin Markov Chain Monte Carlo (LMC) technique is applied, refining the LDCT to align with the Gaussian-noised normal-dose CT distribution. Experiments show superior performance in preserving fine textures and reducing noise compared to conventional methods. Furthermore, as the LMC technique applied, denoising performance of the proposed method is improved in various metrics. This method advances LDCT image processing without requiring paired training data, leveraging the generative capabilities of diffusion models.
13405-93
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Low-dose CT simulation is needed to assess reconstruction/denoising techniques and optimize dose. Image-domain noise-insertion methods face various challenges that affect generalizability, and few have been systematically validated for 3D noise synthesis. To improve generalizability, we presented a physics-informed model-based generative neural network for simulating scanner- and algorithm-specific low-dose CT exams (PALETTE). Using PALETTE, one 2D and two 2.5D models (denoted as 2.5D N-N and N-1) were developed to conduct 2D and effective 3D noise modeling, respectively. These models were trained and tested with an open-access abdominal CT dataset, including 20 testing cases reconstructed with two kernels and various field-of-view. Visual inspection and quantitative assessment were conducted. All models generated realistic CT noise texture. The 2D model provided equivalent or relatively better performance than 2.5D models, showing well-matched local noise levels and high spectral similarity compared to the reference. In summary, PALETTE can provide high-quality low-dose CT simulation to resemble 3D noise characteristics.
13405-94
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
The scarcity of annotated medical images poses a significant challenge to develop and train accurate AI models for the detection of tumors and other pathologies. We introduce a novel method for synthesizing realistic medical images using Multi-Conditioned Denoising Diffusion Probabilistic Models, capable of generating images with and without tumors. Our approach leverages existing datasets to train the diffusion models, ensuring the synthetic images closely mimic real-world medical images. While our primary goal is to increase the volume and diversity of training data, this method also holds potential benefits for underrepresented population groups by facilitating the inclusion of more varied demographic and pathological characteristics. Our results indicate that the diffusion model-generated medical images are indistinguishable from real images by radiologists, demonstrating their potential for effective use in AI model training. We also found that enriching sparse training data with our synthetic images can improve the accuracy of pathology detection AI classifiers. Our paper presents first results on two specific important applications, lung CT and mammography.
13405-95
Compressibility analysis for the differentiable shift-variant filtered backprojection neural network
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This paper presents a novel approach to compress and optimize the differentiable shift-variant FBP neural network based on Principal Component Analysis (PCA).
We apply PCA to the redundancy weights learned from sinusoidal trajectory projection data, revealing significant parameter redundancy in the original model. By integrating PCA directly into the differentiable shift-variant FBP reconstruction pipeline, we develop a method that decomposes the redundancy weight layer parameters into a trainable eigenvector matrix, compressed weights, and a mean vector. This innovative technique achieves a remarkable 97.25% reduction in trainable parameters without compromising reconstruction accuracy.
As a result, our algorithm significantly decreases the complexity of the differentiable shift-variant FBP model and greatly improves training speed. These improvements make the model substantially more practical for real-world applications.
13405-96
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Due to limited photon counts received and various physical degradation factors, further improving PET image quality is essential for quantification and lesion detection. Diffusion models have achieved great success for various inverse-problem tasks, because of their strong representation power and fitness with the Bayesian framework. In this work, we developed a Monte Carlo filtering-based diffusion model for PET image denoising. During the training phase, the score function was pretrained based on normal-dose PET images to learn the prior distribution of PET images. During the inference phase, sequential Monte Carlo posterior sampling (SMC-PS) was applied to achieve accurate estimates of the Bayesian posterior distribution. Also, SMC-PS can balance the trade-off between efficiency and effectiveness, i.e., yielding more accurate estimations with larger particle size. Preliminary experiments based on clinical PET datasets showed that the proposed framework generated better or comparable results than other leading zero-shot or supervised methods for PET image denoising.
13405-97
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Simulation of positron emission tomography (PET) images is generally time- and computationally intensive, primarily involving Monte Carlo or analytical forward projection and iterative image reconstruction. Additionally, since the projection domain typically serves as an intermediary between the object domain and the image domain, the direct relationships between objects and their measurements are rarely studied besides metrics for imaging quality assessment. To address this issue, we develop a novel deep learning-based surrogate model, the noise-aware system generative model (NASGM), for PET image simulation. Specifically, we introduce a generative adversarial network with a novel dual-domain discriminator that processes activity and attenuation maps as inputs and learns various noise characteristics for different acquisition times. Once trained, the network can produce simulated PET images with minimal computational load. Quantitative validation shows that the NASGM generates images with noise levels and distributions most closely matching those of the traditionally simulated images compared to other networks.
13405-98
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Advances in deep learning (DL) provide promising solutions for various aspects of CT, potentially improving image quality and diagnostic outcomes. However, these DL models are known to suffer from performance degradation (i.e., domain shift) when applied to images acquired with unseen imaging protocols (such as different mA, kV, reconstruction kernel) or from different scanner models. In this work, we propose a novel label-free self-supervised learning framework for CT imaging that significantly reduces domain shift by employing a physics-informed contrastive loss (PICL) to train a protocol-invariant encoder (PIE). Compared to the conventional supervised training-from-scratch approach, the proposed approach resulted in ~5.6% improvement in Dice coefficient when tested on the training protocol and ~11.0% improvement in Dice when tested on unseen protocols. These findings highlight the potential of our approach to enhance the reliability and generalizability of deep learning models in CT.
13405-99
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
PET image reconstruction is an inverse problem that aims to reconstruct the radiotracer distribution inside the body. This process involves finding an image 𝑥 that maximizes the likelihood of the observed data 𝑦. To further improve image quality, maximum a posteriori (MAP) reconstruction is developed to exploit prior information about 𝑥, which requires a regularizer. Diffusion models have demonstrated their ability to approximate data distributions. These models traditionally use ancestral sampling to generate images. As research on diffusion models progresses, it has been shown that they can also be used to design regularizers for inverse problems that go beyond ancestral sampling. In this work, we explored diffusion models-based regularizers for PET image reconstruction. Given the computational complexity of directly solving PET reconstruction with such regularizers, half quadratic splitting (HQS) was utilized to decouple the regularizer from the original PET reconstruction objective. Our simulation study demonstrated that the proposed approach could generate high quality reconstructions based solely on a limited number of iterations when compared to other reference methods.
13405-100
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Deep learning-based synthetic CT from MRI could greatly simplify the workflow in MR-based radiotherapy planning and provide more accurate GTV delineation thanks to the better tumor contrast in MRI. Model development for the upper abdomen remains challenging due to the large organ deformation between the CT and MRI. In this work, we proposed a novel distillation method for deep learning-based CT synthesis from MRI. A 3D Swin UNETR was trained and tuned on 192 paired CT and Dixon MR images, followed by postprocessing steps to correct artifacts such as diminishing skins to generate initial synthetic CTs. A distilled Swin UNETR was then trained to map the MRI to the initial synthetic CTs to integrate the postprocessing steps into a single deep neural network. The model was tested on 21 patients with both planning MRI and CT acquired prospectively. The mean absolute error (MAE) between synthetic and planning CT was 62.39 HU. The clinical treatment plans were recomputed on the synthetic CTs. Compared to the planning CT, the synthetic CT achieved a mean PTV dose error of 2.34% and a mean 3%/3mm gamma passing rate of 96.23%.
13405-101
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
The purpose of this work was to evaluate the accuracy of an automated deep learning-based model (nnU-Net) for left ventricular (LV) myocardial segmentation, as compared to semi-automatic LV myocardial segmentation in Vitrea (Canon Medical Systems), where the respective nnU-net and Vitrea segmentations were used for subsequent quantitative CT myocardial rest perfusion measurement in mL/min/g in 40 human subjects. The nnU-Net model demonstrated high accuracy in automated segmentation of the LV myocardium, with a mean Dice score of 0.82. The mean difference in the resultant perfusion measurements was 3.74% (0.0412± 0.0457 mL/min/g), with a Pearson’s correlation of 0.98. Hence, automated nnU-Net LV segmentation may potentially reduce the time and simplify the post-processing necessary for CT myocardial perfusion measurement, without compromising the accuracy of perfusion measurement.
13405-102
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This study aims to enhance the segmentation of heart chambers and myocardium in contrast-enhanced cardiac CT imaging through the optimization of activation and loss functions using the nnU-Net framework. The nnU-Net model was trained using the CHD MICCAI dataset, comprising 68 patients scanned with contrast-enhanced CT. To segment various heart heart chambers and myocardium efficiently, a range of activation functions —CELU, ELU, GELU, LeakyReLU, PRELU, PDELU, SELU, SWISH—and loss functions—Dice+BD, Dice+CE, Dice+Focal, Dice+HD, Dice+IOU, Dice+TopK were integrated and evaluated. Using Dice similarity coefficient to measure a model’s performance, we determined that Swish and GELU activation functions and Dice+TopK and Dice+HD was identified as exceptionally effective. The combination of Swish or GELU activation functions with the DiceTopK or Dice+HD loss function has been found to substantially improve the accuracy and consistency of contrast-enhanced cardiac CT image segmentation.
13405-103
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
In this study, we proposed a multi-agent reinforcement learning (MARL)-based denoising model for restoring digital tomosynthesis (DT) images. Also, the various network architectures for feature extraction were designed in order to evaluate the effect of feature extraction techniques in the MARL-based DT denoising model. The MARL network consisted of the shared, value and policy sub-networks, which implemented feature extraction, reward calculation and optimal policy determination sub-tasks, respectively. Five shared sub-networks were constructed by using convolution layers, dilated convolution layers and residual blocks. The results showed that the proposed models successfully suppressed DT image noise, and the performance of the models with the shared sub-networks including the residual blocks was superior to the other models in terms of denoising, spatial resolution restoration and training efficiency. In conclusion, the modification of the shared sub-network can maximize the performance of the MARL-based DT denoising model.
Posters: Virtual Clinical Trial and Phantoms
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
View poster session description and guidelines above.
13405-105
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Micro-CT imaging in mouse cancer models is vital for developing therapies. Multi-energy CT from a photon-counting detector (PCD) enhances imaging in cancer studies using nanoparticle-based contrast agents or combination therapy. But improving multi-energy CT while limiting radiation dose requires imaging parameter optimization, which is not possible in cancer studies. In silico simulation of photon-counting CT (PCCT) scans of mice with cancer allows parameter tuning to improve image quality. This requires detailed digital mouse phantoms, realistic tumor models, and accurate modeling of imaging effects. We built an in silico PCCT pipeline for mouse cancer studies by using PCCT simulation software, an enhanced mouse whole body (MOBY) phantom, and tumor models from CompuCell3D. We transferred vasculature from a real mouse to MOBY through affine warps. Our PCCT simulation of MOBY with a tumor containing iodine and barium reproduced noise and material contamination seen in real PCCT scans of mice with tumors. We used the simulation input to derive figures of merit for optimizing parameter settings. Future work will refine tumor models and improve simulation accuracy to enhance imaging.
13405-106
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
The extended cardiac-torso (XCAT) phantom, commonly used in medical imaging studies, is a comprehensive computational model that represents human anatomy and physiology. Current XCAT phantoms lack detailed models for coronary artery plaques, limiting their use for cardiac imaging. The purpose of this work was to create a variable Monte Carlo plaque growth algorithm for use with the XCAT phantoms. First, a healthy vessel is created, and the plaque is seeded at a random location inside the vessel wall. At each iteration, probabilities of plaque growth are assigned to each voxel following pathways towards different classes of plaques. Plaques were generated and inserted into an XCAT phantom that was subsequently sent into CT simulation software. The resulting images were visually inspected to determine the conspicuity of the plaques. Results show that the growth algorithm can create a diverse set of plaques that can successfully be inserted into the XCAT phantom.
13405-107
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
The rapid advancement of medical technologies poses significant challenges for researchers and practitioners. In response to the limitations of traditional clinical trials, in-silico trials and digital twins have emerged as transformative technologies, offering efficient and ethical alternatives. In April 2024, Duke University hosted the first international summit on Virtual Imaging Trials in Medicine (VITM), gathering over 130 experts to discuss the latest developments. Key takeaways included the importance of diverse digital patient representations, integrating physics and biology in simulations, and establishing Good Simulation Practices. The summit emphasized the potential of in-silico trials to revolutionize medical research and patient care.
13405-108
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Positron emission tomography-computed tomography (PET-CT) is a frequently used multimodality imaging technique. The combined use of these modalities enables enhanced lesion detection, improved tumor localization, and radiomics analyses. The benefits of dual-modality imaging need to be harnessed using imaging trials. Such imaging experiments can be performed using virtual imaging trials with various advantageous such as low costs and access to ground truth. The purpose of this study was to develop an integrated and validated PET-CT simulation pipeline, enabling virtual imaging trials for dual-modality applications. The developed simulator models both CT and PET acquisition processes, aligns spatial domains to ensure registered geometry between modalities, and reconstructs the PET acquisition data. Verification and validation studies demonstrated anticipated radionuclide uptake regions, properly registered emission and reconstruction data for dual-modality imaging analysis, and close agreement between simulated and experimental measurements. Overall, the pipeline demonstrated to be a viable tool for PET-CT virtual imaging trials.
13405-109
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Medical imaging studies benefit from the use of virtual phantoms since they present a defined ground-truth on the
location and presence of organs. Currently, these phantoms are primarily morphologically focused representations and
thus, do not have accurate information on intra-organ textures, such as heterogeneous material composition. We developed a 3D CT conditional generative adversarial network (3D CTGAN) that solves this problem by synthesizing intra-organ textures for large organ systems such as bone, muscle, and fat. The generated texture phantoms were compared with the original CT scans and showed an improvement of 29% in mean absolute error (MAE), 2% in structural similarity (SSIM), and 7% in peak signal-to-noise ratio (PSNR) compared to current homogeneous composition methods.
13405-110
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
The purpose of this study was to develop an advanced breast suppression technique based on deep learning using a mixed dataset of real and virtual patients for the evaluation of lung function with DCR. Virtual patients with and without breasts were generated by the 4D extended cardiac-torso (XCAT) program. The pix2pix model was trained to estimate the breast-only images from the original images. The resulting breast-only images were subtracted from the original images with breasts to obtain breast-suppression images. The proposed breast suppression technique has the potential to improve the diagnostic performance of pulmonary function evaluation with DCR.
13405-111
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Respiratory motion phantoms (RMPs) are important validation tools for the assessment of respiratory motion compensation technologies including tumor tracking algorithms. Existing lung RMPs are highly simplified in structure, lacking realistic anatomy, attenuation changes, and deformations. However, some technologies such as CT ventilation imaging and deformable image registration rely on nonrigid deformations and attenuation changes observed in patient lungs during respiration, and therefore require RMPs which exhibit these characteristics for testing. To address this issue, we have developed PixelPrint4D, a novel method of 3D printing deformable CT imaging phantoms with realistic anatomic structures and realistic deformation characteristics. This work evaluates the compression behavior of PixelPrint4D phantoms, including attenuation changes and Jacobian determinants, and compares the results to a patient 4DCT. These results contribute to the validation of deformable patient-based phantoms fabricated with PixelPrint4D and will facilitate more robust evaluations of novel CT technologies using PixelPrint4D RMPs.
13405-112
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
In this work, we propose a novel workflow for generating dynamic, patient-specific virtual cardiac geometries to improve the clinical accuracy and relevancy of cardiac virtual clinical trials. Building on our previous image-based approach, we utilize LS-Prepost (licensed by Ansys) to analyze the mid-ventricular diastole frame from 4D cardiac CT data segmented with Simpleware (licensed by Synopsys). These models, while nearly replicating the motion of the previous image based cardiac models, incorporate patient-specific material properties and morphologies. Our finite-element-based method enables the creation of a comprehensive virtual cardiac library, allowing for the manipulation of these patient-specific material and physiological parameters to simulate various cardiac conditions. This advancement is crucial for the future of virtual 4D cardiac imaging and clinical trials.
13405-113
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Post therapy single photon imaging allows for patient dosimetry in radioligand therapy. Monte Carlo simulation enables fast and reliable evaluation of various imaging methods and approaches to inform clinical practice. Its practical use, however, is challenged by the involved intensive computation and the resulting long computation time if performed within CPU. In this work, we aim to develop a GPU-accelerated simulator of single photon imaging and assess its performance. Planar acquisitions were simulated to image both a uniform cylindrical phantom and a Jaszczak phantom. The simulations included both the imaging of 99mTc with a low-energy collimator and and 177Lu with a medium-energy collimator. The results demonstrated that GPU-based Monte Carlo simulation is capable of drastic reduction of computation time while maintaining quantitative accuracy. It allows for quick assessment of imaging methods and approaches for their use in post radioligand therapy single photon imaging and therefore provides a great tool for image based dosimetry research.
13405-114
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Technology has been developed to manufacture custom ultra-high resolution (UHR) computed tomography (CT) phantoms based on patient data using 3D printing. These UHR phantoms replicate lifelike features on sub-millimeter length scales, pushing 3D printed phantoms into resolution regimes that can take full advantage of the UHR capabilities of photon counting CT (PCCT). The methods described exhibit strong robustness and adaptability. When combined with the low costs of 3D printing, this allows for the economical production of custom UHR phantoms. Making these phantoms widely accessible will be vital for research and fully leveraging PCCT scanners' benefits in clinical and educational contexts.
Posters: Detectors
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
View poster session description and guidelines above.
13405-115
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This work compares two metrics to evaluate the improvement in image quality for antiscatter grids employed with energy integrating detectors: (1) the signal-to-noise ratio (SNR) improvement factor (SIF), defined as SIF=Tp/(Tt^0.5), and (2) the detective quantum efficiency improvement factor given by DQEif = SNR^2(grid in)/SNR^2(grid out). Monte Carlo simulations were performed with PENELOPE and penEasy plus the FASTDETECT2 optical transport software. Our results indicate that using SIF slightly overestimates antiscatter grid performance, approximately 2% to 7% compared to the square root of DQEif for the conditions explored in this work.
13405-116
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
We developed an efficient GPU-accelerated Monte Carlo simulation program specialized for X-ray diffraction imaging (XRDI), where molecular interference in coherent scattering takes effect. Besides, multiple non-ideal factors such as charge sharing, charge collection efficiency and ballistic deficit are incorporated in the photon detection and spectrum measurement process. We carried out simulations of pencil beam and coded aperture XRDI to demonstrate the effectiveness of our program. The computation speed shows acceleration of more than 500 times compared to the CPU simulation code.
13405-117
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
The study investigates the impact of detailed semiconductor detector modeling on image quality in photon counting detectors, which offer enhanced contrast, intrinsic spectral imaging, and reduced electronic noise. However, common Monte Carlo codes used in radiation detection simulations often simplify important factors, such as sensor bias and crystalline structure. This study aims to simulate chest images using both the original PENELOPE code and the THOR code extension, which models electron-hole pair (EHP) creation and transport. The simulation setup includes a polyenergetic X-ray source (40-80 kV), a newborn anthropomorphic phantom, a carbon fiber table, and a 1 mm thick CdTe detector biased at -300 V. Image quality was assessed using contrast and signal-to-noise ratio (SNR). The THOR code produced images with 50% fewer counts due to EHP losses, particularly from charge trapping, resulting in lower contrast and SNR compared to the PENELOPE code, with relative differences of up to 6.75% and 39%, respectively. This study presents the impact of a detailed modeling of semiconductor detectors and reveals how neglecting this model leads to an overestimation of image quality.
13405-118
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Most photon counting detectors (PCDs) for CT applications use a direct conversion sensor. Indirect conversion, scintillator-based PCDs have historically been too slow for the high flux requirements of diagnostic CT, but recent fast, bright scintillators inspire us to rethink this paradigm. We evaluate the potential of a LaBr3:Ce PCD using Monte Carlo simulations and show that it may be competitive with CdTe or CZT PCDs.
13405-119
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This study focuses on developing indirect conversion CMOS X-ray detectors as an alternative to traditional TFT-based detectors in medical imaging. Two CMOS indirect X-ray detectors were fabricated utilizing scintillators such as cesium iodide (CsI) and gadolinium oxysulfide. The spatial resolution of the detectors was evaluated using the slanted-edge method to measure the modulation transfer function (MTF). Results show that the CsI-based detector achieved a resolution of 10 lp/mm, while the gadolinium oxysulfide based detector reached 4.5 lp/mm. These values are higher than those of commercial TFT detectors, demonstrating the potential of CMOS detectors to outperform TFT detectors in high-resolution imaging applications. The enhanced spatial resolution suggests that CMOS detectors could be particularly effective for detailed imaging procedures, offering improvements in diagnostic accuracy for various medical imaging applications.
Posters: Novel Imaging Methods
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
View poster session description and guidelines above.
13405-120
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
X-ray computed tomography (CT) is a well-established and frequently used technique in clinical diagnostics due to its non-destructive high-resolution images. There have been several refinements and promising new technologies such as dual-energy CT or photon counting detectors which gain even more detailed insights. All these techniques are restricted to the X-rays’ attenuation contrast only. Dark-field CT, on the contrary, amplifies the radiological information by accessing indirect information on the tissues micro-structure including e.g. porosity by utilizing a grating interferometer.
In this work, we use ventilated ex-vivo porcine lungs to mimic the clinical use case of the technology realistically. We demonstrate that the system has a sufficient sensitivity for lung imaging and obtain a first benchmark for HUd values of lung tissue.
13405-121
Quantification of the dark-field signal and its linear decomposition approximation of packed spheres
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Grating based X-ray dark-field chest imaging using a Talbot-Lau interferometer is a promising technique for the early detection of emphysema and fibrosis. In this study, we focus on the densely packed spheres because the dark-field signals generated by different sizes are similar to those of real lung tissue and we introduce a theoretical model for calculating the dark-field signal of monodisperse packed spheres and use our model to fit experimental data. Then, we examine the dark-field signal of a polydisperse system composed of two different sizes of spheres mixed with equal mass, and explore the theoretical model in the context of polydispersion.
13405-122
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Preliminary prototype has been conducted to evaluate a scalable multiplexed readout electronics for radioactive contaminated seafood imaging system. The radiation detector was composed of a 4x4 CsI(Tl) scintillator and GAPD photosensor array, and may be suitable for fabricating a radioactive contaminated seafood imaging system in the humid conditions of marine environments. Analog and digital signal processing circuit was developed to encode the photon energy and interacted position. Each winner take all (WTA) board reduces 16 analog signals of CsI(Tl)-GAPD radiation detector to a single analog signal and four digital bits that represent the winning channel. A main board was designed to integrate the 16 WTA boards, and to have capacity of multiplexing ratio of 256:1. Analog output signals of radiation detectors were processed in the preamplifier units, offset correction unit and gain adjustment units. Comparator units were used to generate the valid event signal, and FPGA identified the winning channel with fastest signal. Then, the custom-made data acquisition system (DAQ) was utilized to digitize output signals from scalable multiplexed readout
13405-123
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Although optical microscopy provides high-resolution optical images, the light causes blur and the energy decreases exponentially, limiting light penetration depth due to the optical scattering. For this reason, we recently proposed and verified that optrasound-induced deep optical microscopy (OPS-DOM) can provide deep optical imaging by using gas bubble combined optical and ultrasound energies. However, due to restrictions in the size of bubbles generated, there are limitations in selectively acquiring 3D optical images at deep depths. In this study, we proposed a method that controlled the size of the gas bubble by changing the ultrasound frequency and measuring the light penetration depth. Through this, it can be presented to freely and selectively control imaging depth for deep optical microscopy.
13405-124
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
In the era of pandemics, a system that accurately monitors the patient’s condition while preventing infection is needed. In particular, diseases such as COVID-19 require real-time monitoring of the patient’s oxygen saturation, etc. In this study, we propose an automated system that uses video-photoelectric plethysmography (IPPG) to track important health indicators such as oxygen saturation, heart rate, and respiration rate in real time. The system used here uses two wavelengths, 660 nm (red) and 940 nm (NIR). For real-time bio-signal measurement, we focus on improving the accuracy and processing speed of the system by considering appropriate factors such as the region of interest and image acquisition time. To verify the performance of the developed system and software, we measured bio-signals using hairless mice, which have somewhat less light scattering due to hair.
13405-125
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
X-ray machines are essential yet inefficient devices, with conventional systems requiring significant energy to heat filaments and accelerate electrons. Our breakthrough is the development of a palm-sized X-ray system using a Carbon Nanotube (CNT) cold cathode digital X-ray tube, eliminating the filament and simplifying the design. This pocket-sized device weighs just 83 grams, measures 40 mm x 30 mm x 25 mm, and can be operated by AAA and Li-ion batteries. It emits up to 60 mSv/hr of radiation at a 10 cm distance, making it viable for cell irradiation experiments and potentially replacing hazardous materials used in QC testing and calibration of X-ray survey meters. This innovation promises to revolutionize X-ray technology by enhancing portability, energy efficiency, and environmental sustainability.
13405-126
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
X-ray diffraction (XRD) is considered to be a potential non-destructive detection technology in medicine and security inspection. However, the lack of clear knowledge regarding the XRD spectra of prevalent amorphous substances pose impediments to the advancement of this technology. In this work, we scanned over 100 materials to establish a valuable dataset of XRD patterns for further study on XRD phenomena. Using a pencil beam polychromatic XRD system, we conducted scans of thin samples and collected diffraction signals with a pixelated energy-dispersive photon counting detector. XRD patterns were extracted from the signals. The dataset contains XRD spectra of various materials, including crystal powders, tissues, botanical specimens, polymers, geological stones, and common household items. Among them, the XRD patterns of amorphous materials such as water and adipose exhibit excellent congruence with the Geant4 dataset. These patterns serve as strong evidences to our data precision. As an example of applying our dataset, we conducted a simple simulation. In future works, we believe the dataset will facilitate in-depth researches on XRD technology.
13405-127
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Nanotechnology-based digital X-ray sources, such as those utilizing carbon nanotubes (CNT), are revolutionizing the X-ray field. However, direct comparisons between traditional analog filament-based X-ray sources and digital CNT-based X-ray sources remain limited. This study developed X-ray imaging systems incorporating both technologies to evaluate their performance. We found that while the analog filament source produces higher-resolution images in extreme setups due to its smaller focal spot, the CNT source offers superior stability and efficiency considering electrical power over time. The results highlight the CNT source’s advantages for continuous imaging, despite the filament source's superiority in detailed imaging for certain applications.
13405-128
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
The aim of this study was to investigate the accuracy of 225Ac-rituximab liver dosimetry from CT attenuation corrected counts of planar images. Mice were injected with 20 µCi of 225Ac-rituximab. 3D SPECT images were acquired at 2, 24, 48, 72 and 96 hours post-injection then converted to 2D planar images. After converting CT images to attenuation maps, they were applied to the planar images. The activity at each voxel was calculated from geometric mean and CT attenuation corrected count planer image using Python. %ID was obtained from SPECT, geometric mean and CT attenuation corrected planer image and residence time was calculated. S-Values were calculated MC approach. SPECT and planar images exhibited similar liver uptake slopes. Planar imaging showed similar liver uptake values comparing to SPECT. Compared to SPECT-based dosimetry, geometric mean planar images-based dosimetry showed an average difference of 22.7% in the absorbed dose. CT attenuation corrected planar images demonstrated better agreement with SPECT, with only a 12.0% average difference. Our results suggest that combining CT attenuation maps with planar images could be effective alternative to SPECT images.
13405-129
Optimizing transmit field inhomogeneity of parallel RF transmit design in 7T MRI using deep learning
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Ultrahigh field (UHF) Magnetic Resonance Imaging (MRI) offers higher spatial resolution due to an improved signal-to-noise ratio. However, it also presents challenges, such as transmit radio frequency field (B1+) inhomogeneities, causing uneven flip angles and image intensity anomalies. Traditional methods like Magnitude Least Squares (MLS) optimization, although effective, are time-consuming and patient-dependent. Recent machine learning techniques show promise but have limitations. This study proposes a novel, two-step deep learning approach. First, Adaptive Moment Estimation is used to obtain reference RF shimming weights. Then, Residual Networks (ResNets) map B1+ fields to target RF shimming outputs. This method avoids pre-calculated reference optimizations in the testing process, achieving faster and more accurate RF shimming design. Comparative studies with MLS show our method’s superior speed and accuracy, enhancing UHF MRI imaging quality and its medical applications.
Posters: Photon Counting Detector CT
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
View poster session description and guidelines above.
13405-130
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Photon-counting detectors (PCDs) in computed tomography (CT) offer advantages over conventional detectors, such as improved spatial resolution and material discrimination. However, PCDs exhibit greater pixel-to-pixel variability, leading to ring artifacts. This study proposes a calibration method using cylindrical water phantoms to correct these variations. The method involves two steps: acquiring calibration data by positioning phantoms at various offsets from the iso-center and estimating “good pixel” responses to obtain pixel-to-pixel variation correction coefficients. Evaluations using a prototype PCD-based CT scanner showed a significant reduction in ring artifacts and improved image quality in test phantom data, validating the method's effectiveness. This approach enhances PCD-based CT systems and paves the way for future improvements in CT imaging.
13405-131
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
In this simulation study, we investigate the feasibility of photon-counting micro-CT for intraoperative virtual histopathology on tumor specimens, to visualize the tumor margin. We calculate the contrast-to-noise ratio for soft-tissue imaging and investigate its kVp dependence, for single-bin photon-counting mode and optimally energy-weighting mode. Furthermore, we use the Rose model to calculate the minimum detectable feature size as a function of exposure time. To understand the upper limit of detection capabilities, we assume an ideal detector. Our results show that when contrast-to-noise ratio is normalized with exposure time, the maximum studied tube voltage (150 kV) is optimal across tissue thicknesses. In contrast, when CNR is normalized for dose, lower tube voltage ranges are preferable for thinner tissues. Additionally, we find that soft-tissue features as small as 30 μm can be distinguished during a 30-min scan, suggesting that micro-CT may be a feasible alternative to histopathology for intraoperative tumor margin assessment.
13405-132
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This study evaluates the impact of phantom design on iodine quantification accuracy in spectral CT imaging through simulations. Using a photon counting spectral CT system with cadmium telluride (CdTe) detectors, we compared cylindrical (20 cm diameter) and anatomically realistic elliptical phantoms (18/14, 20/16, and 23/18 cm major/minor diameters, 0.5, 0.6, and 0.7 cm skull thicknesses). Each phantom included iodine inserts at concentrations of 0, 2, 5, and 10 mg/ml and diameters of 1, 0.5, and 0.3 cm. The study evaluated the influence of bowtie filters, various tube currents, and operating voltages. Image reconstruction was performed after beam hardening correction using the signal to thickness calibration (STC) method.
The cylindrical phantom, while providing higher accuracy, lacks realism compared to the elliptical phantom with a skull-material layer. The absence of bowtie filters increased discrepancies in root mean square error (RMSE) by up to 3 mg/ml for the highest iodine concentration. This study highlights the importance of selecting appropriate phantom designs for evaluating spectral photon counting head CT systems.
13405-133
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Triple-energy imaging using photon-counting x-ray detectors could enhance real-time background subtraction in interventional radiography. X-ray scatter can impact image quality, so this study aimed to assess scatter levels by energy bin for triple-energy imaging. Using the air-gap method with a 20-cm-thick PMMA phantom and a cadmium telluride photon-counting detector with two energy bins and analog charge summing for charge-sharing suppression, we measured the scatter levels in each energy bin. Data sets for triple-energy imaging were created from two separate dual-energy data sets at tube voltages of 60kV, 80kV, and 100kV, with energy thresholds optimized using contrast-to-noise ratios. The scatter-to-primary ratio (SPR), primary fraction, and scatter fraction were measured across various field sizes. Total SPRs ranged from 0.2 to 2.5, with the SPR varying by up to 2.5 times across energy bins for a fixed tube voltage. The low-energy bin had the highest SPR and the high-energy bin the lowest. These findings are crucial for understanding image quality in triple-energy photon-counting imaging for angiography.
13405-134
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
The purpose of the present work was to assess how NPS properties impact visual perception of low contrast lung nodules across various CT image acquisition protocols. This was achieved by evaluating the quantitative characteristics of the NPS using a previously-proposed NPS parameterization and comparing them to the appearance of a synthetic ground-glass nodule inserted into a lung phantom. This phantom was imaged using an EICT and a PCCT systems. Different dose levels, reconstruction algorithms, and kernels were used. Using similar dose levels in both systems, the use of DLR resulted in improved noise properties and better visualization in comparison to that with a HIR. The improvement of the nodule edge definition using DLR/lung combination is evident in both EICT and PCCT images. Therefore, PCCT/DLR/lung combination demonstrated improved capability to characterize the edges of the low contrast nodule
13405-135
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
The purpose of this work was to evaluate the precision and accuracy of volume measurements of solid nodules (SN) and ground glass opacities (GGO) in images acquired using conventional (EICT) and a photon-counting CT (PCCT) systems, reconstructed using HIR and DLR reconstructions. A 3D printed lung phantom and synthetic nodules were designed and constructed. The ground truth was verified using CT images. CT images of the phantom were acquired with a CTDIvol of 1.4 mGy. Ten acquisitions were performed per scanner/protocol combination with phantom repositioning to mimic clinical positioning variations. The relative error (RE) and the coefficient of variation (CV) were used to evaluate the accuracy and precision of the volume estimates, respectively. The choice of reconstruction method and kernel combination is critical in both systems, affecting both precision and accuracy. When comparing EICT and PCCT systems using the combination DLR/Lung kernel, PCCT demonstrate better accuracy (RE < 14% for SN and < 8.1% for GGO) and precision (CV < 5% for SN).
13405-136
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Energy-resolving photon-counting CT promises improved material-separation capabilities but accurate separation of iodine and calcium remains a challenge. We present proof of concept for a deep-learning based method that takes a pair of basis images from a photon-counting CT and produces a map of the iodine distribution where the contamination from other materials such as calcium is minimized. Ground-truth iodine concentration maps were generated by subtracting 40 keV virtual monoenergetic non-contrast from corresponding contrast-enhanced images after image registration. We then trained a ResUnet++ deep convolutional neural network using water-iodine basis image pairs as inputs and the iodine concentration map obtained from subtraction as label. We demonstrate the performance of this method on clinical images of the carotid arteries and our results show that the trained network correctly highlights the image features containing iodinated contrast agent and quantifies the concentration accurately, with important potential implications for imaging atherosclerosis.
13405-137
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Photon Counting CT often requires high X-ray exposure due to low photon counts per channel, leading to prolonged scanning times, which may not be practical. Here, we explored a trade-off between the number of projections and exposure time per projection to optimize scanning efficiency and image quality. By increasing projections and reducing exposure, we initially generated noisier datasets, which we denoised using unsupervised, data-driven techniques. We applied unsupervised denoising to synthetic spectral CT datasets with a distinct K-edge in the X-ray absorption spectrum. We compared our results with an iterative reconstruction algorithm using 3 regularization parameters (1 spatial and 2 spectral dimensions), which uses fewer projections and higher doses. Although this algorithm employs fewer projections and higher doses, it matches our method in scanning time, allowing a direct evaluation of methods. Our approach significantly reduced scanning time by 36-fold compared to traditional full-dose methods, without compromising image quality. It also eliminates the need for meticulous parameter tuning, simplifying the operational process and enhancing usability.
13405-138
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Photon-counting CT (PCCT) enhances pancreatic cancer imaging, particularly with virtual monoenergetic imaging (VMI) at ≤70 keV, which significantly improves tumor conspicuity compared to conventional CT. PCCT also shows strong inter-reader agreement for assessing peripancreatic vessel involvement and metastasis while reducing radiation dose. Additionally, PCCT has a higher detection rate of pancreatic cysts (4.9% vs. 3% with conventional CT), especially in CT angiograms. Previous studies used vendor-recommended or institutional protocols with standard resolution (SR) mode, but recent findings suggest ultra-high resolution (UHR) mode is preferable. This study evaluates the impact of varying kVp and imaging modes on image quality using a pancreatic phantom. The phantom, mimicking pancreatic parenchyma and tumors, was scanned with PCCT across four kVp and collimation/mode settings. Images were reconstructed as T3D, VMIs at 50, 60, and 70 keV, and iodine map density (IMD). Contrast-to-noise ratios (CNRs) were compared, revealing consistent CT numbers on VMIs and IMD, with UHR mode providing lower noise and higher CNRs, potentially enhancing tumor visibility.
13405-139
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Compared with energy-integrating detectors (EIDs) which use pixelated scintillators, photon-counting detectors (PCDs) employ a continuous semiconductor slab that directly converts each x-ray photon to an electric signal, yielding high detector quantum efficiency. While the benefits of an anti-scatter grid (ASG) have been well investigated for EIDs, the use of ASG on PCDs may be less advantageous due to the reduction in geometric detection efficiency.
Through simulation studies, we compared the influence of various ASG designs, z-collimation widths, and phantom sizes on the SNR for both detector types. Preliminary results indicate that PCDs in low-scatter scenarios may achieve higher SNR with a lower-dimension ASG design, while a 2D ASG consistently enhances the SNR for EIDs. These findings suggest that PCD systems might require different ASG strategies compared to traditional EID systems to maximize performance.
13405-140
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This study evaluates the utility of photon-counting CT (PCCT) spectral localizer radiographs (LRs) in bone density quantification, aiming to enhance opportunistic diagnosis of bone diseases like osteoporosis. Using an anthropomorphic spine phantom, we compared PCCT spectral LRs against the clinical gold standard, dual-energy X-ray absorptiometry (DEXA), with scans performed at 120 kV and 140 kV. Post-processing involved material decomposition to extract water and hydroxyapatite maps, followed by bone segmentation and area bone mineral density (aBMD) calculation. We also explored optimal acquisition and post-processing conditions using a physical phantom and virtual imaging trials. PCCT spectral LRs demonstrated high agreement with DEXA, with percentage differences from 1.7% to 4.5% and a lower coefficient of variation (2.8% vs. 3.1%), highlighting its potential as a reliable tool for early bone disease diagnosis.
13405-141
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This paper introduces a novel framework for evaluating photon-counting computed tomography (PCCT) designs, leveraging neural networks to streamline the comparison of detector and ASIC performances. Traditional methods rely on extensive simulations and the Cramér-Rao lower bound (CRLB) calculations to assess minimum noise, demanding significant computational resources. Our approach simplifies this process by capturing Gaussian-correlated noise characteristics across energy bins in detector pixels, eliminating the need for complex covariance matrix computations. By employing standard simulation datasets representative of photon-counting detectors (PCDs), this method provides an efficient alternative that reduces the computational burden and simplifies the evaluation process. This advancement promises to accelerate the development and optimization of PCCT technologies, making it more feasible to rapidly assess and deploy new designs in clinical applications. This neural network-based approach marks a significant step forward in the practical evaluation of PCCT systems.
13405-142
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This study investigated the impact of bowtie filters on photon-counting detector (PCD) based CT imaging through numerical simulations and benchtop experiments. Results revealed that bowtie filters did not always optimize image quality, with observed effects including capping and flattening. The bowtie filter could influence the uniformity of PCD-CT images based on object size and X-ray spectrum. Adding a thicker beam filter can effectively reduce artifacts, and certain beam hardening correction methods can also effectively address the non-uniformity of PCD-CT images when calibrated properly. These findings highlight the need for careful application of bowtie filters in PCD-CT imaging.
13405-143
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Single photon emission computed tomography (SPECT) traditionally utilizes photon-counting detectors (PCD), advancing PCD systems for CT imaging may have made SPECT/CT with a singular PCD feasible. This has the potential to save costs and reduce co-registration issues between the two systems. Our study aims to determine the upper limits of visibility of a technetium-99m for SPECT/CT using a single detector. Our methodology used TOPAS to simulate an ideal PCD. A 2D-focused collimator was used with a grid ratio ranging from 20:1 to 60:1, and the optimal grid ratio was determined by the modulation transfer function (MTF) and the contrast-to-noise ratio (CNR). Our study confirmed the feasibility of SPECT/CT with a singular photon-counting CT system. We found that a grid ratio 40:1 maximized the CNR at 14.8+/-2 while maintaining a spatial resolution of 0.2 cycles/mm at MTF = 0.1. These findings provide a solid foundation for developing cost-effective imaging systems.
13405-144
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This study evaluates the effectiveness of a Multi-Material Decomposition (MMD) algorithm in Photon-Counting CT (PCCT) for quantifying iodine concentrations and assessing residual calcium errors, essential for accurate cardiac perfusion measurements. Analyzing 50 scans with 10,500 images, the study found that Quantum Iterative Reconstruction (QIR) significantly improves the accuracy and reliability of the MMD algorithm, especially at lower radiation doses. The results also suggest that calcifications have minimal impact on iodine maps, supporting the clinical potential of the MMD algorithm for cardiac perfusion assessment.
13405-145
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Aim: This study aims to evaluate the performance of 3D U-Net segmentation models using various
combinations of Virtual Monoenergetic Images (VMIs) derived from Photon-counting Computed To-
mography (PCCT) and compare their effectiveness against conventional Energy-integrating CT (EICT)
images.
Materials and Methods: The studied cohort comprised 49 PCCT scans and 88 EICT scans of
potential kidney donors. The UNET models were trained with single-channel VMIs at 60keV and
190keV , and dual-channel VMIs combining both energy levels. The performance of these models was
assessed using the Dice Similarity Coefficient (DSC) on PCCT and EICT test sets.
Results and Conclusions: Single-channel models performed well, with the 60keV model achieving a
DSC of 0.92 ± 0.01 on PCCT and 0.87 ± 0.02 on EICT CE, and the 190keV model achieving 0.91 ± 0.01
on PCCT and 0.91±0.02 on EICT NC. The dual-channel model, however, did not significantly improve
performance and demonstrated increased variability, especially on EICT cases. These findings suggest
that while multi-channel inputs hold potential, careful selection and combination of energy levels are
crucial for optimizing segmentation performance
13405-146
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Photon-counting detector (PCD) CT has been introduced for clinical use, and several studies have showcased the benefits of ultra-high-resolution imaging using PCD-CT. Traditionally, CT image spatial resolution is reported near the isocenter where the source-detector alignment and beam geometry are the most optimal. Spatial resolution variability away from isocenter where beam divergence and geometric distortions occur has not been investigated on clinical PCD-CT to the best of our knowledge. To maximize the benefits of UHR PCD-CT and guide the design of optimal imaging protocols, we investigated spatial resolution variability in the clinical PCD-CT at different table heights (0, 5 and 10 cm), reconstruction field of view (50, 100 and 150 mm) and x-ray focal spots (sUHR: 0.4 mm x 0.5 mm, and UHR: 0.6 mm x 0.7 mm focal spot sizes). Modulation transfer function (MTF) showed a percent decrease (10% MTF) up to 11.8% off-isocenter (10 cm table height offset relative to scan isocenter). A 12.6% decrease in 10% MTF value was observed between sUHR and UHR focal spots.
13405-147
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Clinical photon-counting computed tomography (PCCT), FDA-approved in 2021, enhances image quality and diagnostic performance with better dose efficiency than conventional CT. PCCT offers ultra-high-resolution (UHR) and standard-resolution (SR) imaging. Studies suggest UHR images provide lower noise and potential dose reduction due to the small pixel effect. However, differences in noise power spectrum (NPS) and low-contrast detectability (LCD) between UHR and SR are underexplored. This study evaluates the small pixel effect on NPS and LCD using a standard ACR CT phantom scanned on a clinical PCCT (NAEOTOM Alpha, Siemens). Results show significant noise reduction with decreased radiation dose and increased kernel sharpness, but minimal change with slice thickness. UHR mode demonstrates superior low-contrast detectability and overall image quality across all parameters, suppressing more medium and high spatial frequency noise and consistently improving image quality.
13405-148
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Contrast enhancement is a key feature in detecting breast malignancies, making CT a valuable tool for breast cancer imaging. This study evaluated the spectral performance of photon-counting CT (PCCT) using phantoms. A mammography phantom with simulated breast lesions (iodine concentrations 0.2–2.0 mg/mL) was placed atop a chest phantom with matching iodine rods. Imaging was performed on a clinical PCCT scanner (NAEOTOM Alpha, Siemens) in standard-resolution (SR) and ultra-high-resolution (UHR) modes. Images were reconstructed using the Qr44 kernel and Quantum Iterative Reconstruction (QIR-3) with large (410 mm) and small (185 mm) fields of view (FOV) and matrix sizes of 512 and 1024. Various image types, including threshold-low, virtual monoenergetic images (VMIs) at 50, 60, and 70 keV, and iodine maps, were generated. The analysis of circular regions of interest (ROIs) within the breast lesions showed minimal impact from FOV and matrix size on quantification and lesion circularity. UHR mode consistently yielded lower normalized bias, noise, and root mean square error (RMSE), with 50 keV VMIs providing the best spectral performance for breast tumor imaging.
13405-149
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Photon-counting computed tomography (PCCT) enhances clinical diagnostics by accurately measuring the energy of X-ray photons, facilitating the differentiation of materials and tissues. PCCT utilizes projections from multiple energy bins for material decomposition—a challenging task due to nonlinearities and physical non-idealities like charge-sharing and pulse pile-up in detector and ASIC responses. Material decomposition algorithms are categorized into image-domain, projection-domain, and one-step inversion decomposition. The latter, though time-consuming, integrates decomposition and reconstruction into a joint optimization framework, offering a comprehensive solution that avoids spectral information loss and reduces artifacts compared to other methods. Recently, Implicit Neural Representation (INR) has been applied to this complex inverse problem, establishing a direct mapping between spatial distributions and material attenuation, supported by a physics-guided deep learning model that enhances accuracy. Preliminary results using a prototype PCCT system with water and Gammex phantoms demonstrate the effective one-step material decomposition capability of this approach.
13405-150
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
The majority of acute myocardial infarcts (AMI) occur in non-calcified plaques with <50% stenosis. Low-attenuation plaque (LAP) is this most important compositional plaque feature predicting future development of AMI. A recent societal expert consensus statement concluded that accuracy and reproducibility of LAP quantitation by CT is poor. Photon counting CT (PCD-CT) has shown ultra-high resolution (UHR) acquisition mode reduces blooming and partial volume averaging (PVA) artifacts in comparison to standard resolution mode with helical full rotation CT angiography (CTA), however its impact on LAP detection is unknown. Therefore, we evaluated the impact of UHR on detection of LAP in a mixed coronary plaque (-60 HU pericardial fat, 1000 HU iodine lumen, 75 HU LAP, and 300 HU calcified plaque (CP)) with a clinical coronary CTA acquisition with simulated gating in comparison to ground truth. We show a profound impact of iodine and CP on LAP detection in both UHR and standard reconstructions. These limitations can provide a realistic foundation upon which to optimize PCD-CT CCTA protocols.
Posters: Breast Imaging
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
View poster session description and guidelines above.
13405-151
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
The goal of this project is to develop a convolutional neural network (CNN) that can automatically score the
CDMAM phantom placed adjacent to the SunNuclear (formerly CIRS) swirl slab and blocks of PMMA. This
CNN will be released to the public for evaluation of DBT systems.
Our initial findings suggest that the pre-trained baseline model can be used as a starting point for the CNN-
based observer, and then cross-validation runs can be carried out to evaluate a particular system’s performance
with new data. The assumption, that was verified on the systems tested herein, is that performance in the
cross-validation runs will saturate at a maximum value.
13405-152
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This work focuses on implementing a Monte Carlo (MC) model to simulate photon-counting (PC) detectors, using a previously validated Geant4-based MC code. MC simulations of bCT acquisitions, using both PC and EI detection technologies, were performed to investigate image quality improvement in projection and reconstructed images based on signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR). The PC and EI detectors were composed of CdTe and CsI, respectively. The evaluation of PC simulation performance for a 50 keV mono-energetic beam, considering a 20 keV energy threshold and four energy bins, shows that the detection efficiency was 88%, with 24% of the counts corresponding to the incorrect energy channel. For image quality evaluation, 225 projection images for bCT were simulated using anthropomorphic breast phantoms with an added iodinated mass and a W/Cu (0.257 mm) at 65 kV x-ray spectrum. CT images were reconstructed using FPB. For the projections, the SNR and CNR values increased, respectively, by up to 88/95% and 31/35% for single/multichannel PC vs. EI-based detectors. For the reconstructed images, the improvements were up to 35% and 5%.
13405-153
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
In this study, we propose a deep learning based two-stage breast CT reconstruction in the image domain. We evaluated the proposed method on the AAPM 2021 sparse view CT reconstruction challenge dataset. The experimental results demonstrate that the proposed method performs better than all comparison methods.
13405-154
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
The generation of synthesized mammographic images is pivotal for advancing virtual clinical trials and enhancing machine learning models through data augmentation. In this study, it is presented a comprehensive framework that integrates ray-tracing methods, Monte Carlo simulations, and deep learning for scatter estimation to efficiently generate synthesized mammographic images. Anthropomorphic breast phantoms were used to generate a dataset of 6220 simulated images across different beam energies, which were used to train a deep learning model based on the ResNet architecture. The model achieved a mean percentage error of 2.49% in scatter prediction compared to MC simulations. Overall, the complete pipeline is reliable to generate synthesized images, with mean percentage differences of 3.74% compared to MC images, on average.This approach significantly reduced computational time by an order of magnitude compared to traditional MC methods, enabling a faster generation of synthetic images, compared to traditional methods. The proposed pipeline facilitates the creation of large, diverse datasets, supporting the optimisation of image analysis and virtual clinical trials in mammography.
13405-155
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Despite advances in imaging and histology, discrepancies and sampling errors persist, prompting research into x-ray diffraction methods like Wide Angle X-ray Scatter for breast cancer diagnosis. This study estimates the differential linear x-ray scattering coefficient (mu_s) of stroma and uses a three-basis function method to estimate the mu_s of breast specimens. The basis functions are mu_s coefficients of stroma, epithelial cells, and fat. The fractional volumes for stroma, cells, and fat denoted by nu_str, nu_cell, nu_fat, respectively, are calculated using the singular value decomposition method. For a cancer specimen, nu_str= 0.58, nu_cell = 0.43, nu_fat = -0.01, whereas for a benign one nu_str= 0.94, nu_cell = 0.11, nu_fat = -0.05. The negative volume coefficients for fat indicate its absence from the samples. Cancerous samples exhibit higher cell content, proving uncontrolled cell-proliferation is a cancer hallmark, while benign samples primarily reflect increased “fibrous” tissue. The fitting algorithm is effective for characterizing composite breast specimens, with results aligning well with known cancer biology.
13405-156
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Scatter radiation in X-ray breast imaging degrades contrast, hindering lesion detectability. As scatter signals are primarily low-frequency, downsampling is commonly applied in deep-learning-based correction methods. We investigated scatter estimation networks across image domains and downsampling ratios using VICTRE and MC-GPU simulations. Networks were trained in the inverted intensity and log-transformed domains with downsampling ratios of 2x2, 8x8, and 16x16. Performance was evaluated on scatter-corrected images restored to the original resolution. A U-Net model with residual connections was trained using mean absolute error (MAE). Mean squared error (MSE) and structural similarity index (SSIM) were calculated. Log-transformed training resulted in higher pixel accuracy (lower MSE), while the inverted intensity domain achieved higher SSIM, indicating better structural preservation. We suggest log-transformed domain training for tasks requiring pixel accuracy, whereas the inverted intensity domain is more suitable for preserving details, which is crucial for detecting anomalies in X-ray breast imaging.
13405-157
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
L2 represents a breast phantom with a structured, non-static background. By shaking the phantom between image acquisitions, different background configurations are achieved. This study assesses the variability of the background across multiple models of the phantom.
Four models of the phantom where manufactured. 60 2D image acquisitions of one model and each 12 2D image acquisitions of the three further models where performed at automatic exposure control (AEC) dose level. Regions of interest were defined in the non-static background of the phantom models, followed by radiomic feature extraction using the open-source software Lifex (www.lifexsoft.org).
Out of 110 extracted radiomic features, 16 were found to be highly reproducible across all image acquisitions. The variability within the 4 models of the phantom was found to be within the range of variability observed in a single model. This indicates that the manufacturing process of the non-static background is reproducible with respect to the extracted features.
Posters: CT Image Quality
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
View poster session description and guidelines above.
13405-158
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Static CT is a CT system geometry that has a source ring and a detector ring and hence is not composed of any rotating components. One of the major issues of systems with a detector ring is scattered radiation as the deployment of anti scatter grids is not feasible. This issue may be overcome with new methods such as deep scatter estimation (DSE). This work investigates the use of DSE for static CT. Our results show that DSE allows to correct very well for scatter artifacts while outperforming a kernel-based reference method. The performance of DSE can be further improved by using three projections as network input to provide additional depth information of the object.
13405-159
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
The goal of this work is to study non-unique solutions in dual-energy CT (DECT) using an example object containing gadolinium, which has a k-edge in the diagnostic x-ray range. Fluence measurements through the object using a 70 kVp and 120 kVp x-ray spectrum are simulated and material pathlengths are determined using contour intersections and optimization. For noiseless transmission data, the contour method recovers the true pathlengths of the object and a second solution pair at regions in the object. For noisy transmission data, solving using the optimization method over 1000 realizations reveals an estimate of the pathlength bias and variance. The non-uniqueness of the the estimated pathlengths in the noiseless case corresponds to excessive bias when estimating the material pathlengths in the noisy case.
13405-160
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Chronic obstructive pulmonary disease (COPD), encompassing chronic bronchitis and emphysema, requires accurate CT imaging for reliable assessment. However, inconsistent imaging protocols often compromise measurement reliability. This study aims to optimize CT protocols for cross-sectional and longitudinal COPD measurements using a virtual imaging framework. We modeled various stages of emphysema and bronchitis conditions and simulated a clinical CT scanner. The impact of focal spot size, tube current, and kernel sharpness on COPD biomarkers, LAA-950 (emphysema) and Pi10 (bronchitis), were analyzed. Results showed that increasing tube current reduced variability in LAA-950 but minimally affected Pi10, while sharper kernels increased variability and error for both LAA-950 and Pi10, highlighting the importance of standardized protocols for accurate and consistent COPD assessment.
13405-161
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
The increase of commercially available Computed Tomography (CT) scanner technologies has expanded patient access to advanced imaging technologies but has also introduced potential sources of variability that could impact patient care. This study investigated the influence of intra- and inter-scanner variability on image quality and quantitative imaging tasks, which was accomplished through the evaluation of detectability index (d’). We analyzed 631 clinical phantom images from the COPDGene study and assessed image quality metrics: Noise Power Spectrum (NPS) and Modulation Transfer Function (MTF). Results demonstrated observable variations in NPS and MTF across identical acquisitions, and notable differences in d’ values across different scanner models. These findings emphasize the importance of considering scanner variability in patient care and encourage the development of clinically-diverse virtual CT scanner models to better gauge its impact.
13405-162
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Accurately assessing health of the lung parenchyma is critically important in the management of patients with
chronic obstructive pulmonary disease (COPD). CT attenuation values have been shown to have clinical value
for this purpose. However, CT attenuation values lack absolute accuracy due to various physical effects. In this
work, we conduct an experimental study to evaluate, realistically and in isolation, the effect of beam hardening
errors. Our initial results indicate that spectral interactions between the rib cage and the vertebra induce complex
beam hardening error patterns within the lung parenchyma that vary from patient to patient as well as with the
slice location. For a difference of 10 HU between healthy and diseased tissue, these variations cause a drop of
0.06 in discrimination performance measured by the area under the ROC curve.
13405-163
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
In this work, we propose a metal artifact reduction method using implicit neural representation and dual-domain regularization. Our approach utilizes neural network implicit priors, total variation regularization in the image domain, and linear regularization in the sinogram domain. On a widely used simulation dataset, our method achieves superior quantitative metrics and visual results, outperforming traditional and some supervised and unsupervised learning methods that need extensive training data. Importantly, our method works case by case and requires no training data.
13405-164
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Metal Artifact Reduction (MAR) in Computed Tomography (CT) is crucial for clinical diagnosis and treatment planning, as metallic implants often cause severe artifacts like streaking and shading, compromising image accuracy. These artifacts arise from issues like beam hardening, photon starvation, and scatter Recently, diffusion models have shown promise in addressing inverse problems in medical imaging. This paper proposes a dual-domain MAR approach using a score-based diffusion model. We first train a diffusion model on a large dataset of artifact-free CT images to learn the distribution of clean images. During sampling, we apply the diffusion prior in the sinogram domain, using Poisson blending to fill metal trace regions. This is followed by image fusion in the image domain to refine the images further. To improve efficiency, we use the Come-Closer-Diffuse-Faster (CCDF) algorithm, which accelerates the process by starting with an MAR with Linear interpolation (LMAR) image that has been enhanced with noise corresponding to a specific sampling step.
13405-165
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
In CT imaging, balancing radiation dose and image quality is crucial for effective clinical diagnosis. Traditionally, physicians or medical physicists had used noise power spectrum on a phantom to measure the noise level in CT image. However, this study introduces a novel approach for automated NPS measurement in clinical CT data using deep learning-based organ segmentation and structure coherence feature techniques. In phantom datasets, as the strength of IR method increases, it was observed that the spatial frequency peak shifts towards lower frequencies in both Siemens and Canon scanners. The peak shift in spatial frequency was observed to be 0.077 cycles/mm for Siemens and 0.0434 cycles/mm for Canon. Clinical data from 210 patients demonstrated variability in the peak noise magnitude between individuals, representing low consistency between individuals. This highlights that one-fits-all strategy using phantom-based dose optimization may not be appropriate for clinical utility and emphasizes the need for patient-specific image quality optimization. We expect that proposed method could contribute to improving CT image quality and radiation safety in clinical diagnosis.
13405-166
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
In recent years, a lot of novel spectral filters has been proposed, such as split filter, spectral modulator, and dynamic bowtie filter. However, due to the inherent size of the focal spot, these filters characterized by rapid spatial variations in X-ray attenuation, can cause spectral mixing in the penumbra area of the detector. To address this challenge, we establish a multi-ray model propose an adaptive subsampled weighting of filter thickness (A-SWIFT) method, which approximates the mixing spectra in the penumbra region by weighted the spectra attenuated by multiple filters. Our simulation and experiment results with a spectral modulator show that the ring artifacts in reconstructed images after beam-hardening correction can be well suppressed by the A-SWIFT method; Simulations with a split filter also show that the spectra at the center penumbra region can be better estimated by A-SWIFT method than the traditional method.
13405-167
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Enhancing the spatial resolution of CT images is a relentless pursuit in the field of computed tomography (CT). Two common strategies for achieving this goal are the use of advanced hardwares or an super-resolution (SR) algorithms. While using advanced hardware can improve spatial resolution, it also significantly increases system costs. according to the data processing inequality and previous research, improving diagnostic information through deep learning-based SR in the image domain is challenging. In this paper, we propose a hybrid approach that combines both hardware and algorithmic improvements to enhance the spatial resolution CT reconstruction. Denoising diffusion probabilistic model (DDPM) is employed to achieve SR reconstruction. During the reverse diffusion process, the HR projection prior is integrated into repeated projection and reconstruction operations.
13405-168
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
In this study, the application of NIQE and BRISQUE to assess the clinical CT image quality was investigated the LDCTIQAC2023 dataset, while adapting their scoring system. Our approach involves refining NIQE to focus the method on relevant anatomical areas. This adaptation significantly improves the algorithm score, from 2.38 to 2.70, with 3.00 being the maximum score achievable. BRISQUE achieved a score of 2.78 without modification. While BRISQUE required a larger dataset of 800 quality-labelled images, 87 best-scoring images were sufficient for our refined NIQE method (NIQE-CT). Our results highlight the potential of NIQE-CT and BRISQUE as effective tools for developing reliable and efficient image quality assessment models for clinical CT scans, with the aim of optimising diagnostic accuracy while ensuring patient safety through effective radiation dose management.
13405-169
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
This study investigates how energy thresholds impact material decomposition bias in photon-counting computed tomography (PCCT). Using a projection-based model with a Shepp-Logan phantom, we examined how varying energy thresholds from 20 to 90 keV affect the accuracy of decomposing images into brain/bone and brain/iodine materials. Results showed that lower thresholds (<40 keV) increased decomposition bias, with significant peaks near iodine’s k-edge. Bias decreased with higher thresholds (>50 keV), particularly for non-basis materials. This trend was consistent across different virtual monoenergetic images at 60 keV and 140 keV. The findings highlight the critical role of energy thresholds in PCCT accuracy.
13405-170
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Time-resolved imaging, such as 4DCT, is essential for tracking tumors in motion due to breathing. As multiple volumes are acquired, each volume contains increased noise due to lower radiation doses. Denoising these scans is beneficial, and 3D trainable bilateral filtering, which requires minimal data and provides explainable results, is a promising method. This study extends this filtering into the temporal domain for application on 4DCT and shows strong denoising performance and image content reconstruction with improved preservation of small structures in the lung. Expert ratings indicate substantial agreement and a general preference for the 4D filtering method in most cases.
13405-171
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
We introduce a method to obtain tomographic uncertainty images in X-ray computed tomography (XCT). The method called Stochastic Variational Tomography (SVT), uses stochastic variational inference and treats the tomographic reconstruction process as a Bayesian inverse problem. As a result, SVT produces computationally efficient uncertainty reconstructions and integrates smoothly with existing reconstruction pipelines. Notably, we implement the method using the ASTRA tool kit, with no changes to the underlying code. By using in-silico and in-vitro XCT datasets, we are extending previous work, which applied SVT to ultrasound data. In addition, we compare the uncertainty reconstruction with the absolute error, with results that suggest an application to no-reference image quality assessment.
13405-172
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
To enhance the consistency of care and unbiasedly inform optimization design in computed tomography, it is essential to understand why radiation dose and image quality data exhibit variability across similar attenuating body sizes. In this study, we deployed virtual imaging trial techniques to assess whether such variability can be induced by CT scanner capability to effectively adapt x-ray output to different patient positions and body habitus. Three anthropomorphic computational XCAT phantoms were imaged using the DukeSim simulation platform following vertical shifts and different arms positions. Specifically, CTDIvol and average Global Noise Index, organ noise magnitude, and Water Equivalent Diameter were evaluated and compared for a total of 18 simulated imaging conditions. The study highlighted and quantified the effect of patient mispositioning on radiation dose and image quality variability, with arms positions playing a dominant role.
13405-173
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Triple-source CT (TSCT) has potential to offer improved temporal resolution for cardiac imaging but introduces complex cross-scatter interactions. This study presents a novel scatter correction method for TSCT, combining a one-dimensional anti-scatter grid (1D ASG) with beam blockers. Utilizing a custom-built TSCT imaging system with three X-ray tubes and three detectors arranged azimuthally, we investigated the feasibility of scatter correction for TSCT with a focused 1D ASG and beam blockers. Experimental results using a contrast-to-noise ratio (CNR) phantom shows that our method can effectively remove both cross-scattering and forward scattering artifacts, improving CNR by 40% compared to uncorrected images.
13405-174
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
Classic computed tomography reconstructs the attenuation field in a discrete manner. This preset discretization limits resolution, leads to model mismatch, and induces forward-backward projection inconsistency. We propose a transmission ray-bundle functor combined with implicit neural field to capture X-ray tomography. The forward projection operates as a hybrid mapping from continuous attenuation field to the discrete detector units. This setup circumvents the need for explicit reconstruction, ray-voxel tracing, or footprint analysis, and supports high-quality rendering with arbitrary resolution. We evaluated its superiority in both contrast and noise behaviors with modulation transfer and noise power spectrum assessment. Furthermore, performance was quantitatively verified across a wide range of projection numbers. Our method achieves an average PSNR of 29.78 dB for volume reconstruction at 512×512×110 resolution using 20 projections, comparable to the quality of FDK, PWLS-TV, and NAF with 360 views.
13405-175
17 February 2025 • 5:30 PM - 7:00 PM PST | Golden State Ballroom
Show Abstract +
High-resolution microCT scanners require accurate calibration of their physical projection model. The high magnification used to attain high-resolution amplifies mechanical imperfections that makes projection models using constant parameters inadequate to represent a moving and changing system and unsuitable for high quality image reconstruction. To address these problems, we have developed: (i) a dynamic projection model whose parameters vary as a function of gantry angle, and (ii) a customized recalibration adapted to every scan. The dynamic model and custom recalibration produced images of constant quality, able to compensate for changing – and non-reproduceable – environmental conditions. The optimization step that calibrates the model operates globally and is self-calibrating.
Tuesday Morning Keynotes
18 February 2025 • 8:30 AM - 10:00 AM PST | Town & Country A
View Full Details: spie.org/tuesday-morning-keynotes
8:30 AM - 8:35 AM:
Welcome and introduction
8:35 AM - 8:40 AM:
Award announcements
- Robert F. Wagner Award finalists for conferences 13406 and 13412
- Image Processing Student Paper Award
13406-504
Integrating computational approaches to unravel alzheimer's disease and co-pathologies: a biomarker-driven approach to precision medicine
(Keynote Presentation)
18 February 2025 • 8:40 AM - 9:20 AM PST | Town & Country A
Show Abstract +
Early detection and intervention in neurodegenerative diseases hold the potential to significantly impact patient outcomes. This presentation will explore the development of multi-disciplinary and multi-modality biomarkers to identify individuals at risk and monitor disease progression. By combining advanced imaging techniques, such as MRI, and PET, with fluid biomarkers, we aim to detect subtle changes in brain structure and function that precede clinical symptoms. These biomarkers could serve as powerful tools for early diagnosis, enabling timely intervention and potentially delaying disease onset. Furthermore, by identifying individuals at highest risk, we can optimize the design of clinical trials and accelerate the development of effective therapies. Ultimately, our goal is to improve the lives of individuals with neurodegenerative diseases through early detection, precise diagnosis, and targeted treatment.
13412-505
Wearable ultrasound technology
(Keynote Presentation)
18 February 2025 • 9:20 AM - 10:00 AM PST | Town & Country A
Show Abstract +
The use of wearable electronic devices that can acquire vital signs from the human body noninvasively and continuously is a significant trend for healthcare. The combination of materials design and advanced microfabrication techniques enables the integration of various components and devices onto a wearable platform, resulting in functional systems with minimal limitations on the human body. Physiological signals from deep tissues are particularly valuable as they have a stronger and faster correlation with the internal events within the body compared to signals obtained from the surface of the skin. In this presentation, I will demonstrate a soft ultrasonic technology that can noninvasively and continuously acquire dynamic information about deep tissues and central organs. I will also showcase examples of this technology's use in recording blood pressure and flow waveforms in central vessels, monitoring cardiac chamber activities, and measuring core body temperatures. The soft ultrasonic technology presented represents a platform with vast potential for applications in consumer electronics, defense medicine, and clinical practices.
Break
Coffee Break 10:00 AM - 10:30 AM
Session 4:
Photon Counting Detector CT
18 February 2025 • 10:30 AM - 12:30 PM PST | Town & Country B
13405-16
18 February 2025 • 10:30 AM - 10:50 AM PST | Town & Country B
Show Abstract +
For PCCT systems, maximum-likelihood estimation (MLE) offers asymptotically unbiased and efficient (minimum variance) material decomposition but is usually solved iteratively, which is computationally expensive and time-consuming. Conversely, empirical methods relying on grid calibration aim to construct a direct measurement-decomposition conversion, which can be fast but may suffer from bias or noise amplification. In this work, we show that the iterative MLE method implicitly defines the functional mapping from measurements to decomposition. The corresponding mean and noise yield analytical approximation forms from the Implicit Function Theorem. From this perspective, we demonstrate that it is possible to distill knowledge from the implicit function defined by the iterative MLE, i.e., finding the explicit proxy, by leveraging universal approximators and the derivative-aware Sobolev Learning paradigm. We show that the proposed method, namely proxy MD, is both computationally efficient (providing >300 times speedup) and approaches the performance of iterative MLE, thus outperforming conventional empirical methods and enabling high-quality real-time quantitative spectral imaging.
13405-17
18 February 2025 • 10:50 AM - 11:10 AM PST | Town & Country B
Show Abstract +
Photon counting detector (PCD)-based computed tomography (CT) faces challenges at extremely low photon counts, leading to increased bias and noise in image reconstruction. This study introduces a novel framework that optimizes the trade-off between bias and noise in post-log data. Computer simulations show that the proposed method achieves 55.83% reduction in bias and 63.86% reduction in variance within the sinogram domain on average compared to other methods implemented in this study. It also reduces root mean square error (RMSE) by 35.0% and enhances contrast-to-noise ratio (CNR) by 62.7% on average for the reconstructed images. These improvements indicated the proposed method’s ability to enhance the clinical viability of low-dose PCD-CT, and support more accurate quantification and diagnosis.
13405-18
18 February 2025 • 11:10 AM - 11:30 AM PST | Town & Country B
Show Abstract +
Photon counting detector (PCD) CT scanners are associated with ring artifacts that stem from pixel inhomogeneities. Because of their smaller pixel size, these scanners are resolution- limited by the focal spot: concentrating more energy onto a smaller area would melt the focal track. We propose a hybrid 3rd generation/4th generation architecture that could decrease ring artifact by virtue of its sampling geometry, and could also utilize thermal energy more efficiently at the focal track, so that the focal spot could be shrunk while retaining the same power.
13405-19
18 February 2025 • 11:30 AM - 11:50 AM PST | Town & Country B
Show Abstract +
In addition to the objects or patients that are scanned in a CT, other components in the beam path can also cause scatter. In particular, scatter by the bowtie filter, which is typically the last filter before the patient, can create image artifacts, impair contrast and considerably distort the CT values. Previous studies have shown that deep convolutional neural networks are very effective correcting scattered photons in clinical CTs. In this work, we present an adapted approach of the deep scatter estimation (DSE) method for photon-counting (PC) CT that effectively corrects both object and bowtie scatter. Our best network reduces the mean absolute error of object and bowtie scatter from 8.1 HU to 0.9 HU.
13405-20
18 February 2025 • 11:50 AM - 12:10 PM PST | Town & Country B
Show Abstract +
Non-linear partial volume (NLPV) effect in energy-resolving detectors was previously demonstrated in the projection-domain to improve spatial resolution by leveraging spectral information. In this work, we further explore the NLPV algorithm and demonstrate its utility in improving the spatial resolution in PCCT images with simulated phantom studies. We improve sub-pixel iodine profile characterization in the NLPV algorithm to generate the NLPV-enhanced iodine projections and reconstruct iodine images. PCCT scans of phantoms pertinent to knee anatomy were simulated, and we found that with the improved NLPV algorithm, the image-domain spatial resolution improved by ~64% in the noiseless case and ~40% in the noisy case.
13405-21
18 February 2025 • 12:10 PM - 12:30 PM PST | Town & Country B
Show Abstract +
Ultra-high-resolution (UHR) photon-counting detector (PCD) CT offers superior spatial resolution compared to conventional CT, benefiting various clinical areas. However, the UHR resolution also significantly increases image noise, which can limit its clinical adoption in areas such as cardiac CT. In clinical practice, this image noise varies substantially across imaging conditions, such as different diagnostic tasks, patient characteristics (e.g., size), scan protocols, and image reconstruction settings. To address these challenges and provide the full potential of PCD-CT for optimal clinical performance, a convolutional neural network (CNN) denoising algorithm was developed, optimized, and tailored to each specific set of conditions. The algorithm's effectiveness in reducing noise and its impact on coronary artery stenosis quantification across different patient size categories (small: water equivalent diameter <300 mm, medium: 300-320 mm, and large: >320 mm) were objectively assessed. Reconstruction kernels at different sharpness, from Bv60 to Bv76, were investigated to determine optimal settings for each patient size regarding image quality and stenosis assessment.
Break
Lunch Break 12:30 PM - 1:50 PM
Session 5:
Breast Imaging
18 February 2025 • 1:50 PM - 3:10 PM PST | Town & Country B
13405-22
18 February 2025 • 1:50 PM - 2:10 PM PST | Town & Country B
Show Abstract +
This paper introduces a novel approach to quantitative contrast-enhanced spectral mammography (CESM)
using triple energy K-edge imaging and a two-pass material decomposition method. The primary objective
is to accurately determine local projected iodine concentration by leveraging the iodine K-edge in presence of
glandular and adipose breast tissue. Addressing the ill-posed problem of established dual-energy methods, our
approach accounts for three base materials with three non-redundant measurements. The proposed method
involves two key steps. First, we estimate the local total thickness D via three material decompostion, followed
by strong denoising. Second, we use D as a constraint comnputing the recombined iodinge image in classical
dual-energy technique. Initial simulation studies using a numerical CIRS dual-energy phantom demonstrate
effective background cancellation, accurate reconstruction of relative projected iodine concentrations at moderate
image noise levels. In conclusion, the proposed method provides a robust theoretical framework for accurately
reconstructing projected iodine concentration images in CESM
13405-23
18 February 2025 • 2:10 PM - 2:30 PM PST | Town & Country B
Show Abstract +
Direct-indirect dual-layer flat-panel-detector (DLFPD) is a promising technology for dual-energy digital breast tomosynthesis (DEDBT). Scatter correction and breast thickness map estimation are essential tasks to fully realize the clinical advantages of DLFPD. We proposed a multi-task training approach using a single convolutional neural network for simultaneous estimation of scatter maps and breast thickness map for DLFPD-based DEDBT. The trained network accurately estimated scatter maps and central breast thickness map, while faithfully capturing the thickness roll-off in the peripheral region. In future work, we will optimize the network, extend it to MLO views, and eventually apply it to patient images. It is expected to benefit iodine quantification and volumetric breast density calculation for DLFPD-based DEDBT.
13405-24
18 February 2025 • 2:30 PM - 2:50 PM PST | Town & Country B
Show Abstract +
Simulation of temporal changes in breast tissue has been a challenging task, despite significant development in virtual clinical trials. This study aims to enable virtual trials over time. We present a framework for simulating 3D voxel phantoms and breast lesions, using our algorithm based on Perlin Noise, in three modules. In the Population Creator, users can simulate cases based on real population characteristics. In Phantom Creator and Lesion Creator, users select the breast and lesion shape, size, and density. Perlin Noise parameters are selected to match the appearance of different tissue types. We used open-source software to project and reconstruct phantom DBT images. Assuming the volumetric breast density of 10.7% at 57 years and an exponential decrease over time, we simulate the anatomy at the inclusion and exclusion in the screening program (40 and 74 years in Sweden). The breast density was calculated to be 16.0% and 7.2%, respectively. Similarly, we simulate lesions at different times, assuming a doubling time of 282 days. We present a framework for simulating temporal changes in breast tissue, to support the use of virtual trials over time.
13405-25
18 February 2025 • 2:50 PM - 3:10 PM PST | Town & Country B
Show Abstract +
Four-dimensional dynamic contrast-enhanced dedicated breast CT (4D DCE-bCT) is a novel imaging technique designed to characterize tumor heterogeneity by imaging iodinated contrast agent wash-in and wash-out throughout breast tumors. This study describes the development and testing of 3D-printed tumor perfusion phantoms for the validation of 4D DCE-bCT. The tumor phantoms have different designs, consisting of small channels, leaky vessels, and gyroid structures. 4D DCE-bCT images revealed the expected wash-in and wash-out at the phantom entrance and exit, intensity differences between channel sizes, and contrast pooling in the leaking vessel phantom.
Break
Coffee Break 3:10 PM - 3:40 PM
Session 6:
Physics/Image-Guided Procedures: Joint Session with Conferences 13405 and 13408
18 February 2025 • 3:40 PM - 5:20 PM PST | Town & Country B
Session Chair:
Shuai Leng, Mayo Clinic (United States)
13405-26
18 February 2025 • 3:40 PM - 4:00 PM PST | Town & Country B
Show Abstract +
We introduce a novel learning-based approach for rigid patient motion estimation in interventional C-arm CBCT. Several motion compensation strategies exist for different domains of the imaging process, including reconstruction-free data consistency assessments, using epipolar consistency conditions, and reconstruction-based autofocus methods, optimizing image quality metrics. Consistency-based approaches are merely sensitive to out-of-plane motion, while autofocus methods can detect in-plane motion but tend to converge into local minima for larger motion amplitudes. We aim to leverage the strengths of both methods by combining them in a deep learning-based model. Our model consists of two parallel feature extraction networks followed by a classifier, and is trained and tested on simulated motion trajectories applied to motion-free clinical acquisitions. The network predicts 6 rigid directional displacement probabilities for each of the 496 projections. First tests achieve a good classification performance with an average ROC-AUC of 0.9527 and PR-AUC of 0.3961.
13408-23
18 February 2025 • 4:00 PM - 4:20 PM PST | Town & Country B
Show Abstract +
Transcatheter arterial embolization procedures have become more complex and widely used, but the lack of quantitative measures for embolization endpoints limits standardization, relying heavily on the experience of interventional radiologists. In this work we use a benchtop flow system and a 3D printed phantom of hepatic arterial vasculature to build an in vitro model of the dynamic conditions of the blood flow proximal to the catheter tip location during transcatheter embolization of the hepatic arteries. By controlling flow rates and using optical imaging, reproducible vessel occlusion conditions are simulated. The derived image-based metrics, such as time constants and area under the curve of the time-dependent image intensity profile, show consistent patterns and a strong correlation with occlusion levels, suggesting their potential as quantitative measures for embolization endpoints.
13405-27
18 February 2025 • 4:20 PM - 4:40 PM PST | Town & Country B
Show Abstract +
Quantitative digital subtraction angiography (qDSA) is a method for measuring blood velocity from 2D x-ray sequences in interventional radiology. Previous studies in swine models and phantoms compared qDSA to MRI or ultrasound. This study reports a computational fluid dynamics (CFD) simulation platform for controlled investigations of qDSA behavior. Iodine injections into arterial blood flow were simulated with OpenFOAM for different catheter geometries (0-45° angle, 0-180° rotation) and blood velocities. qDSA was applied to projections derived from CFD results. qDSA was linearly related to downstream CFD velocity and catheter orientation changes resulted in 5% standard deviation in qDSA velocity estimates.
13408-24
18 February 2025 • 4:40 PM - 5:00 PM PST | Town & Country B
Show Abstract +
In this work, we developed a deep CNN-based spectral metal artifact reduction algorithm for interventional oncology applications (sMARIO) and quantitatively assessed its performance, with a focus on iodine detection. Patient data was used to train the algorithm, which was subsequently applied to phantom images containing water and iodinated rods adjacent to metallic ablation applicators and used to assess its performance. Spectral MARIO effectively corrected substantial contributions of metal artifacts in the image data across all keVs investigated (40-150 keV). The resulting performance of sMARIO on 70 keV images and below (AUC=0.98) demonstrates the ability to accurately classify small iodine concentrations even in close proximity of metallic objects used in IO procedures.
13405-28
18 February 2025 • 5:00 PM - 5:20 PM PST | Town & Country B
Show Abstract +
We have developed an algorithm called IPEN (intra-intervention perfusion assessment using a standard x-ray system with no gantry rotation) to allow for perfusion assessment in interventional (IR) suites. IPEN puts this 4-dimensional (4-D = 3-D plus time) image reconstruction problem into a well-posed problem framework such that there exists a robust and rigorous solution. We have validated the accuracy with brain and liver perfusion applications. Recently we have further developed IPEN to address patient motion problems, as motion artifacts are common in clinical settings. The method called MC–IPEN uses native (i.e., non-subtracted) x-ray angiography (XA) images and compensates for the effect of head motion very effectively. The use of XA images, however, could be a problem for many hospitals, because they save digital subtraction angiography (DSA) images routinely but not XA images. In this work, we will further develop MC–IPEN such that it works with DSA images and does not need XA images.
Wednesday Morning Keynotes
19 February 2025 • 8:30 AM - 10:00 AM PST | Town & Country A
View Full Details: spie.org/wednesday-morning-keynotes
8:30 AM - 8:35 AM:
Welcome and introduction
8:35 AM - 8:40 AM:
Award announcements
- Robert F. Wagner Award finalists for conferences 13408 and 13413
- Early-Career Investigator Award: Image-Guided Procedures, Robotic Interventions, and Modeling
- Student Paper Award: Image-Guided Procedures, Robotic Interventions, and Modeling
13408-506
Image-guided surgery and examinations using ultrasound
(Keynote Presentation)
19 February 2025 • 8:40 AM - 9:20 AM PST | Town & Country A
Show Abstract +
Many of today’s cancer surgeries are carried out with robot assistance. Using real-time intra-operative ultrasound, we can overlay pre-operative imaging into the surgeon’s console, enabling visualization of sub-surface anatomy and cancer at the same time with the standard laparoscopic camera view. We will discuss aspects of system design, visualization and registration methods that enable such visualization, and present our results. We will also present tissue and instrument tracking approaches that can be used in future augmented reality systems.
For remote and underserved communities, we developed a teleultrasound approach that relies upon using a novice – the patient, a family member or friend – as a robot to carry out the examination. The novice wears a mixed reality headset and follows a rendered virtual ultrasound transducer with the actual transducer. The virtual transducer is controlled by an expert, who sees the remote ultrasound images and feels the transducer forces. This tightly-coupled expert-novice approach has advantages relative to both conventional and robotic teleultrasound. We discuss our system implementation and results.
13413-507
The future of diagnostics: the role of computational pathology in tomorrow's medicine
(Keynote Presentation)
19 February 2025 • 9:20 AM - 10:00 AM PST | Town & Country A
Show Abstract +
Computational Pathology has already led to remarkable innovations in diagnostics, achieving expert pathologist performance in tasks such as prostate cancer grading and cancer metastasis detection. In recent years, we have seen rapid advances, with weakly supervised models able to predict patient outcomes or genetic mutations and foundation models enabling application to rarer diseases. However, this only scratches the surface of what will be possible in the near future.
In this talk, I will briefly touch on the history of computational pathology and how we got to where we are today. Subsequently, I will highlight the current methodological innovations in the field and their potential for causing a paradigm shift in diagnostic pathology. I will discuss how these innovations, combined with the AI-driven integration of radiology, pathology, and 'omics data streams, could change the future of diagnostics as a whole. Last, I will discuss the challenges and pitfalls moving forward and how we, as a community, can contribute to addressing them.
Break
Coffee Break 10:00 AM - 10:30 AM
Session 7:
Image Reconstruction with Diffusion Models
19 February 2025 • 10:30 AM - 12:30 PM PST | Town & Country B
13405-29
19 February 2025 • 10:30 AM - 10:50 AM PST | Town & Country B
Show Abstract +
This research primarily aims to address the reconstruction problem of limited-angle CT. In this work, we utilized a generative artificial intelligence model to complete the limited sinogram, resulting in a full sinogram that can be used for reconstructing medical images. During the sinogram completion process, we also applied the physical properties of the sinogram as constraints to enhance the precision of the completion.
13405-30
19 February 2025 • 10:50 AM - 11:10 AM PST | Town & Country B
Show Abstract +
Dynamic PET can quantitatively recover physiology-related parameters and is increasingly used in both research and clinical settings. However, reconstructing high-quality parametric images from dynamic PET data remains challenging due to the ill-posed nature of kinetic-model fitting and limited data counts. In this work, we proposed a diffusion model-based kinetic-model fitting framework utilizing the Patlak model, which pre-trains the score function on static total-body PET images and uses it as a prior for Patlak slope and intercept images generation. Preliminary results based on total-body dynamic datasets demonstrated promising performance.
13405-31
19 February 2025 • 11:10 AM - 11:30 AM PST | Town & Country B
Show Abstract +
X-ray scatter degrades image quality in computed tomography, particularly in cone-beam CT due to wide cone
angles. While mono-energetic CT scatter correction is well-studied, spectral CT imaging presents additional
challenges due to its sensitivity to unmodeled biases in material decomposition and density estimation. This work
presents a joint estimation approach that simultaneously estimates scatter and material densities by integrating
the scatter component into a spectral CT forward model. Using Diffusion Posterior Sampling method, we leverage
the combination of prior knowledge from large dataset training and the physical model for joint density and scatter
estimation. Tested on simulated and phantom data, our method significantly reduce artifacts associated with
unestimated scatter, improving spectral CT image quality.
13405-32
19 February 2025 • 11:30 AM - 11:50 AM PST | Town & Country B
Show Abstract +
Neurological Positron Emission Tomography (PET) is a critical imaging modality for diagnosing and studying neurodegenerative diseases like Alzheimer’s disease. However, the inherent low spatial resolution of PET images poses significant challenges in clinical settings. This work introduces a novel Generative Super-Resolution (GSR) approach using Fourier Diffusion Models (FDMs) to enhance the spatial resolution of PET images. Unlike traditional methods, FDMs leverage the time-dependent Modulation Transfer Function (MTF) and Noise Power Spectrum (NPS) to generate high-resolution, low-noise images from low-resolution inputs. Our method was evaluated using simulated data derived from High-Resolution Research Tomograph (HRRT) PET images with 2 mm resolution. The results demonstrate that FDMs significantly outperform existing techniques, including conditional diffusion models and image-to-image Schrodinger bridge, across several metrics, including structural similarity and noise suppression. Our simulation results highlight the potential of FDMs to generate high-quality 2mm resolution reconstructions given 4mm resolution input PET data.
13405-33
19 February 2025 • 11:50 AM - 12:10 PM PST | Town & Country B
Show Abstract +
In this research, we propose strategies for 3D DPS CT reconstruction using a 3D neural network to learn the prior distribution. We develop modifications to a standard DPS algorithm to substantially reduce memory requirements and to accelerate the sampling speed. We evaluate different alternatives that permit 3D DPS in realistic CT volume sizes and compare relative merits of each strategy.
13405-34
19 February 2025 • 12:10 PM - 12:30 PM PST | Town & Country B
Show Abstract +
This study investigates bias and precision in conditional Denoising Diffusion Probabilistic Models (DDPMs) for CT imaging. Using the LIDC-LDRI chest CT dataset, we analyze both image- and sinogram-domain models under ideal conditioning. Our findings reveal notable bias and variance, particularly in the sinogram domain, where conventional U-Net architectures struggle to capture long-range features. Even with ideal conditioning, uncertainty may remain too high for certain CT image interpretation tasks. These results underscore the need for improved models and additional physical constraints to mitigate hallucinations and enhance the precision of conditional diffusion models in medical imaging.
Break
Lunch Break 12:30 PM - 1:50 PM
Session 8:
Angiography and Radiography
19 February 2025 • 1:50 PM - 3:10 PM PST | Town & Country B
13405-35
19 February 2025 • 1:50 PM - 2:10 PM PST | Town & Country B
Show Abstract +
Scatter reduces the accuracy of dual-energy (DE) material estimates. However, its impact likely depends on the condition of decomposition, which in turn depends on energy separation of the DE system. We investigate this tradeoff for various realistic DE radiography setups. The study involved Monte Carlo simulations of multi-material abdomen phantoms presenting a range of vertebral CaHA densities. Polyenergetic DE radiographs were obtained in a clinically standard PA contact-scan setting. Four DE protocols at matched doses were investigated: one single-exposure multi-layer detector and three kV-switching protocols. Projection decomposition with denoising was applied to estimate lumbar areal bone mineral densities (aBMDs). Scatter correction with tunable residual errors was emulated by subtracting a fraction of true scatter from the total measured projections. Accuracy of aBMD estimates was compared for different DE protocols and scatter errors. Results show that narrower spectral separation required more accuracy correction in general. Interestingly, accurate aBMD could be obtained at certain of LE-to-HE scatter error ratios, even if the residual scatter error is high.
13405-36
19 February 2025 • 2:10 PM - 2:30 PM PST | Town & Country B
Show Abstract +
Digital subtraction angiography (DSA) is the gold standard imaging technique for neurological and extremity imaging, but is not feasible for cardiac applications due to significant tissue motion between the mask and contrast image. Using energy subtraction angiography (ESA), a set of tissue suppressed images can be acquired rapidly, leaving only bone and iodine visible in the image. In a phantom study using an anthropomorphic chest phantom, a custom 3D printed iodine vessel was shifted on the chest to simulate cardiac motion, and a set of consecutive ESA images were acquired. By evaluating the motion of the vessel across the ESA image frames, a synthetic mask image was created using an algorithm to evaluate the change in attenuation in individual pixels across the image set. Once subtracted from an ESA image, a DSA-like image of just the vessel remained. Additionally, a stationary attenuation filter was used to appeal to potential clinical translation, and a noise reduction algorithm is employed compensate for the noise increase as a result of the subtraction.
13405-37
19 February 2025 • 2:30 PM - 2:50 PM PST | Town & Country B
Show Abstract +
X-ray fluoroscopy is utilized for real-time device guidance during vascular interventions. Deep learning denoising offers the potential to reduce image noise in real-time during these interventions. One approach to train a network for fluoroscopy denoising is to use digital angiography (DA) images as targets and add noise to simulate fluoroscopy. However, the added noise needs to account for variable x-ray beam quality. This work proposes utilizing a noise model that includes variable beam quality and does not require prior knowledge of the DA acquisition parameters. Performance of neural networks trained with a variable beam quality noise model and a noise model from a single beam quality were compared. The network trained with a noise model that included variable beam quality had significantly higher structural similarity index measure (p<0.001), signal-to-noise-ratios (p<0.001), and less noise (p<0.001). The developed method offers a fluoroscopy denoising algorithm training that is insensitive to variable beam quality.
Show Abstract +
The permitted input power density of rotating anode X-ray sources restricts the spatial X-ray image resolution. However, computed tomography would benefit from much smaller focal spots with equal output. Our group is proposing new tungsten microparticle targets that promise an order of magnitude improvement of the focal spot input power density, thus tripling the MTF limit.
Before investing, the limitations of classical technology should be known. Therefore, we modeled target erosion of rotating anodes that allows us to compute a new criticality parameter. In this context we also suggest a new correction factor for calculations of the patient X-ray dose. In conclusion, the specifications of X-ray tubes are justified. Unfortunately, the gain with increasing tube voltage is smaller than predicted by some volume heating models. Tungsten/rhenium coated carbon fiber reinforced anodes promise a few dozen percentage points MTF improvement but cannot compete with newly proposed non-eroding microparticle-based targets.
Break
Coffee Break 3:10 PM - 3:30 PM
Session 9:
Virtual Clinical Trials
19 February 2025 • 3:30 PM - 5:30 PM PST | Town & Country B
13405-39
19 February 2025 • 3:30 PM - 3:50 PM PST | Town & Country B
Show Abstract +
The four-dimensional extended cardiac-torso (XCAT) phantom is widely used in virtual imaging trials (VITs) due to its realistic computational model of human anatomy. However, the current XCAT phantom has limitations representing the beating heart, such as the absence of a detailed model of the extended coronary vasculature and restricted capabilities to simulate disease. This study aims to address these limitations by applying a modern tree generation algorithm to extend the coronary vasculature within XCAT heart models, and integrating this with models for motion, coronary plaque generation and insertion, and blood flow to create a framework to simulate diseased states of the heart. The utility of this framework was demonstrated through an example CT simulation comparing plaque visualization with standard and photon-counting CT. The results illustrate the significant potential of these tools to simulate both normal and abnormal cases to evaluate cardiac imaging techniques.
13405-40
19 February 2025 • 3:50 PM - 4:10 PM PST | Town & Country B
Show Abstract +
Most of the major CT systems in the United States have been cleared with a low-dose CT (LDCT) lung cancer screening (LCS) option for their FBP and iterative image reconstruction methods. Nowadays, deep learning (DL)-based reconstruction and denoising (DLR/D) methods are becoming widely available. There is a need to assess whether DLR/D can preserve diagnostic image quality for various patient characteristics and pathology status in LDCT LCS. These assessments are likely to require a huge amount of testing data, which may not be possible with patient scans. In this work, we explore the feasibility of utilizing a deep virtual CT workflow for these assessments. Preliminary results based on 2D simulations are presented to demonstrate the potential utility of this workflow for evaluating DLR/D in LDCT LCS.
13405-41
19 February 2025 • 4:10 PM - 4:30 PM PST | Town & Country B
Show Abstract +
A scatter-window and DL-based AC method for MPI SPECT (CTLESS) was observed promising performance on the cardiac defect-detection task in a single-scanner setting. To assess its clinical potential, it is important to evaluate the generalizability of CTLESS across different SPECT scanners. In this study, we conducted an in silico imaging trial, namely ISIT-GEN, to assess the generalizability of CTLESS across scanners from three vendors. We evaluated the performance of CTLESS on data acquired from 266 virtual patients who underwent 99mTc-sestamibi one-day stress MPI studies protocol on a Philips Precedence SPECT/CT system, a GE Discovery 670 Pro SPECT/CT system, and a Siemens Symbia Evo Excel SPECT/CT system. The performance was evaluated on the defect-detection task using an anthropomorphic model observer. We observed that CTLESS generalized well across all scanners, consistently yielded statistically non-inferior performance compared to the standard-of-care CT-based AC method, and significantly outperformed a non-AC method. Results of ISIT-GEN also highlight the value of using in silico imaging trials to conduct rigorous and feasible generalizability evaluations.
13405-42
19 February 2025 • 4:30 PM - 4:50 PM PST | Town & Country B
Show Abstract +
We propose a hybrid method of inserting simulated microcalcification clusters into patient projection images acquired on a prototype breast CT scanner with an average mean glandular dose of 6.0 mGy. Ray tracing was used to generate projection images of clusters of five microcalcifications which varied in diameter, chemical composition, and density. These simulated projection images were added to patient projection images and reconstructed using the Feldkamp filtered backprojection algorithm with varying apodization filters. Volumes of interest (VOIs) and maximum intensity projections (MIPs) were extracted from the reconstructed volumes, and deep learning model observers and ROC curve analysis were used to evaluate cluster detectability in a binary classification setting. This methodology may be useful for assessing breast CT systems and other x-ray-based imaging technologies across a broader set of parameters.
13405-43
19 February 2025 • 4:50 PM - 5:10 PM PST | Town & Country B
Show Abstract +
In recent years, computed tomography (CT) imaging require high-resolution reconstructions that can reveal more details of patients.
However, noise from X-ray and blurring effect caused by detector binning disrupts acquiring high-resolution images.
These limitations can be addressed using image restoration techniques, and in particular, implicit neural representation-based techniques that allow flexible adjust output image resolution have been proposed in the field of low-level computer vision.
In this work, we propose a method to remove noise and blurring artifacts while performing continuous CT image reconstruction with consuming less memory and time.
Our method allows radiologists to modify the resolution of the selected region of interest (ROI) image in order to help the diagnosis of more detailed parts.
13405-104
19 February 2025 • 5:10 PM - 5:30 PM PST | Town & Country B
Show Abstract +
In medical imaging, harmonization plays a crucial role in reducing variability arising from diverse imaging devices and protocols. Patient images obtained by different scan conditions of computed tomography may show varying performance when processed by the same artificial intelligence model or quantitative metrics. The purpose of this study was to explore and analyze the role of harmonization within a virtual imaging trial platform. An evaluation framework consisting of three typical task-based scenarios was proposed: lung structure segmentation, chronic obstructive pulmonary disease (COPD) quantification, and lung nodule quantification. Evaluation results before and after harmonization reveal three findings: 1) Improved Dice scores and reduced Hausdorff Distances at 95th Percentile in lung structure segmentation; 2) Decreased variation in biomarkers and radiomics features in COPD quantification; and 3) Increased number of features with high intraclass correlation coefficient in lung nodule quantification. The results demonstrate the significant potential of harmonization across various task-based scenarios.
Thursday Morning Keynotes
20 February 2025 • 8:30 AM - 10:00 AM PST | Town & Country A
View Full Details: spie.org/thursday-morning-keynotes
8:30 AM - 8:35 AM:
Welcome and introduction
8:35 AM - 8:40 AM:
Award announcements
- Robert F. Wagner Award finalists for conferences 13407 and 13410
- Computer-Aided Diagnosis Best Paper Award
13407-508
Pioneering vision: The journey of building AI for transformative healthcare
(Keynote Presentation)
20 February 2025 • 8:40 AM - 9:20 AM PST | Town & Country A
13410-509
Machine learning in neuroimaging: Understanding heterogeneity of neurologic diseases and building personalized imaging-AI biomarkers
(Keynote Presentation)
20 February 2025 • 9:20 AM - 10:00 AM PST | Town & Country A
Show Abstract +
Machine learning has transformed medical imaging in general, and neuroimaging in particular, during the past two decades. We review our work in this field, starting with early contributions on developing personalized predictive markers of brain change in aging and Alzheimer’s Disease, and moving to recent weakly-supervised deep learning methods, aiming to dissect heterogeneity of brain change in neurodegenerative and neuropsychiatric diseases, as well as in brain cancer. We show that disease-related brain changes can follow multiple trajectories and patterns, which have distinct clinical and genetic correlates, thereby suggesting a dimensional approach to capturing brain phenotypes, using machine learning methods.
Break
Coffee Break 10:00 AM - 10:30 AM
Session 10:
CT Image Quality
20 February 2025 • 10:30 AM - 12:30 PM PST | Town & Country B
13405-44
20 February 2025 • 10:30 AM - 10:50 AM PST | Town & Country B
Show Abstract +
Spectral CT thermometry presents a unique opportunity to improve local tumor control with intraprocedural temperature monitoring of thermal ablations. Previous evaluation of spectral CT thermometry in ex vivo bovine liver and liver-mimicking phantoms demonstrated feasibility and met clinical requirements. Ablation of in vivo porcine liver revealed a strong relationship between physical density and temperature that corresponded to thermal volumetric expansion. This validation of spectral CT thermometry in vivo portends its clinical translation and potential to reduce local tumor recurrences for thermal ablations.
13405-45
20 February 2025 • 10:50 AM - 11:10 AM PST | Town & Country B
Show Abstract +
A convolutional neural network was trained for denoising high-resolution (HR) energy-integrating detector (EID) cardiac CT images. To supervise the training, noise-only images were estimated from patient images using two different methods: Subtraction of low and high iterative reconstruction (IR) strength reconstructions; subtraction of adjacent image slices with the same IR strength. After training, the models were applied to an independent cohort of HR cardiac EID-CT images.
Models trained with either noise-image generation method reduced noise in HR EID-CT images by 74-79%. However, the model trained on subtraction of low and high IR images resulted in undesirable salt-and-pepper noise texture and CT number bias. The model trained on adjacent slice subtraction images had more natural texture and no significant bias. Both models preserved spatial resolution of the inputs and did not alter anatomical structures.
13405-46
20 February 2025 • 11:10 AM - 11:30 AM PST | Town & Country B
Show Abstract +
This study presents the development, implementation, and testing of a 3D printed phantom with two compartments designed for dynamic micro-CT imaging using low molecular weight contrast agents. The phantom was evaluated through both optical and micro-CT imaging to assess its ability to generate and repeat various time attenuation curves (TACs). This preclinical phantom promises to be a valuable tool for validating and quantifying perfusion micro-CT measurements. This work represents one of the first adaptations and implementations of a dynamic perfusion phantom for CT at the preclinical level, providing a standardized method for quality assurance in preclinical and research settings.
13405-47
20 February 2025 • 11:30 AM - 11:50 AM PST | Town & Country B
Show Abstract +
Task-based image quality assessment is essential for CT protocol and radiation dose optimization. Despite many ongoing efforts, there is still an unmet need to measure and monitor the quality of images acquired from each patient exam. In this work, we developed a patient-specific channelized Hotelling observer (CHO)-based method to estimate the lesion detectability for each individual patient scan. The ensemble of background was created from patient images to include both relatively uniform regions and anatomically varying regions. Signals were modelled from lesions of different sizes and contrast levels after incorporating the effect of contrast-dependent spatial resolution. Index of detectability (d’) was estimated using a CHO framework. This method was applied to clinical patient images from 3 different CT scanner models and at 3 different radiation dose levels. The d’ for 5 different lesion size/contrast conditions was calculated across the scan range of each patient exam. The average noise level and d’ were 13.1/3.68, 16.8/2.93 and 21.7/2.35 for a patient scanned at 100%, 50% and 25% dose levels, respectively.
13405-48
20 February 2025 • 11:50 AM - 12:10 PM PST | Town & Country B
Show Abstract +
Cardiac motion presents a significant challenge in maintaining image quality in cardiac CT scans. Despite advances in temporal resolution and optimized scan protocols, heart movement during scans can still cause severe motion artifacts, reducing the diagnostic value. A common solution involves partial angle reconstructions (PARs), where the scan is divided into angular intervals, each reconstructed separately. When the number of PARs is chosen correctly, each PAR represents a quasi-static motion state. Motion between these PARs is corrected using motion vector fields (MVFs), mapping them from their respective motion states to a reference state before combining them for a final, artifact-free image. Unlike methods focused only on coronary arteries, this approach extends to the entire heart. A residual U-net, trained predicts the aforementioned MVFs. In tests, this method significantly reduced motion artifacts across various heart regions, decreasing the average mean absolute error (MAE) from 32.74 HU to 8.37 HU.
13405-49
20 February 2025 • 12:10 PM - 12:30 PM PST | Town & Country B
Show Abstract +
Virtual monoenergetic images (VMIs) from photon counting detector CT (PCD-CT) provide distinct clinical benefits. Lower keV VMIs enhance iodine and bone contrasts but struggle with blooming artifacts, while higher keV VMIs effectively reduce beam hardening, blooming, and metal artifacts but diminish contrast, presenting a trade-off among different keV levels. To address this, we introduce a contrast-guided virtual monoenergetic image synthesis framework (CITRINE) utilizing adversarial learning to synthesize images by integrating beneficial spectral characteristics from various keV levels.The synthesized images showed reduced blooming artifacts, similar to those observed at 100 keV VMI, and exhibited high iodine contrast in the coronary lumen, comparable to that of 70 keV VMI. Notably, compared to the original 70 keV VMI, CITRINE images achieved approximately 25% reduction in percent diameter stenosis while maintaining consistent contrast levels. These results confirm CITRINE’s effectiveness in improving diagnostic accuracy and efficiency in cCTA by leveraging the full potential of multi-energy and PCD-CT technologies.
Break
Lunch Break 12:30 PM - 1:50 PM
Session 11:
Phase Contrast and Dark Field Imaging
20 February 2025 • 1:50 PM - 3:10 PM PST | Town & Country B
13405-50
20 February 2025 • 1:50 PM - 2:10 PM PST | Town & Country B
Show Abstract +
Existing diagnostic techniques, including imaging, often struggle to detect early lung changes effectively. However, considering x-rays as electromagnetic waves opens up additional contrast mechanisms, such as diffraction, phase-shift, and small-angle scattering. In healthy pulmonary alveoli, x-ray scattering produces a strong darkfield signal, which decreases when alveolar integrity is compromised. X-ray darkfield imaging shows promise for in-vivo lung evaluation but is currently unavailable for clinical and preclinical use due to complex, expensive, and shock-sensitive hardware requirements. Speckle-based darkfield imaging addresses these challenges, generating darkfield contrast images with simpler equipment. This study introduces an in-vivo small animal lung darkfield x-ray imaging setup, utilizing a liquid-metal jet x-ray source and a photon-counting detector. The results validate the system's ability to measure microstructures, highlighting its potential for preclinical in-vivo lung assessment.
13405-51
20 February 2025 • 2:10 PM - 2:30 PM PST | Town & Country B
Show Abstract +
Grating-based dark-field and phase x-ray imaging aims to extend conventional attenuation-based x-ray imaging by revealing additional microstructural information. A big challenge for translating grating-based dark-field computed tomography to medical applications lies in minimizing the data acquisition time. While a continuously moving detector is ideal, it prohibits conventional phase retrieval and stepping techniques that require multiple projections under the same angle with different grating positions. In this work, we introduce a new algorithm that improves phase retrieval for continuously acquired data and mitigates movement-based cross-talk artifacts.
13405-52
20 February 2025 • 2:30 PM - 2:50 PM PST | Town & Country B
Show Abstract +
The X-ray dark-field signal of an object varies depending on its position along the ray path. Consequently, in dark-field CT, the signal of every given voxel in an object is not constant during its rotation, leading to inaccuracies with conventional reconstruction methods. This work proposed an analytical reconstruction formula that addresses the positional dependent problem by treating the dark-field CT as a weighted Radon transform and applying a corresponding inversion formula. This reconstruction formula was validated through both numerical simulated and experimental data.
13405-53
20 February 2025 • 2:50 PM - 3:10 PM PST | Town & Country B
Show Abstract +
X-ray phase contrast and dark field imaging offer significant potential to overcome the limitations of traditional absorption-based imaging by enhancing contrast for soft tissues and reducing radiation doses. These advanced imaging techniques respectively leverage phase and small-angle X-ray scattering (SAXS) information to provide superior visualization of soft materials and unresolved micro-structures. In this study, we introduce an innovative benchtop multi-contrast imaging and computed tomography (CT) system capable of simultaneously capturing absorption, differential phase, phase, and dark-field images in a single shot per projection angle. Our innovative imaging setup addresses common challenges in the field by eliminating the need for a highly coherent X-ray source and/or an ultra-high-resolution detector, or X-ray gratings with very small periods or high aspect ratios, thereby simplifying the implementation and reducing associated costs. The proposed system's versatility and efficiency make it a promising tool for various biomedical applications, offering a significant advancement in X-ray imaging technology.
Break
Coffee Break 3:10 PM - 3:40 PM
Session 12:
Deep Learning Applied to Imaging Physics
20 February 2025 • 3:40 PM - 5:20 PM PST | Town & Country B
13405-54
20 February 2025 • 3:40 PM - 4:00 PM PST | Town & Country B
Show Abstract +
Medical implants made of dense materials pose great challenges to accurate CT reconstruction. This work proposes a novel approach for joint anatomy and implant estimation using a mixed prior model. This prior model leverages a learning-based diffusion prior for anatomy and a simple 0-norm sparsity prior for implant. Additionally, a hybrid mono-polychromatic forward model with free parameters is employed to effectively accommodate the spectral effects of implants. The whole reconstruction process alternates between two subproblems: diffusion posterior sampling is used to update the anatomy image, and classic optimization updates the implant image and spectral coefficients. Evaluation on spine imaging with screw implants shows that the proposed algorithm can achieve accurate anatomy reconstruction between two screws. Besides the spine region, proposed algorithms also effectively avoided streaking and beam hardening artifacts on soft tissue, achieving 15.25% higher PSNR and 24.29% higher SSIM compared to normalized metal artifacts reduction (NMAR). Simulation results demonstrated promising performance of the proposed algorithm on accurate reconstruction of anatomy and implant.
13405-55
20 February 2025 • 4:00 PM - 4:20 PM PST | Town & Country B
Show Abstract +
This study investigates the use of a convolutional neural network to perform micro-CT perfusion quantification. The ability to quantify perfusion metrics such as blood flow, blood volume, and mean transit time provides valuable insights into tissue viability and function, aiding in diagnosis, treatment planning, and monitoring therapeutic responses. Preclinical micro-CT perfusion imaging holds significant promise for advancing our understanding of various physiological and pathological processes in small animal models. Various methods have been developed to quantify perfusion metrics from CT data; however, these methods have notable drawbacks, particularly their voxel-by-voxel nature which introduces significant noise and variability into the perfusion maps. In this work, we demonstrate a potential deep learning approach to perfusion quantification. The network input consisted of 20 timepoints along random geometric arrangements of typical time attenuation curves. The output of the network consisted of 4 parametric maps representing the numerical parameters of a gamma variate curve. This approach led to reduction in noise and accurate recreation of time attenuation curves.
13405-56
20 February 2025 • 4:20 PM - 4:40 PM PST | Town & Country B
Show Abstract +
Aggressive radiation dose reduction and the intrinsic uncertainty in convolutional neural network (CNN) outputs are detrimental to detecting critical lesions (e.g., liver metastases) in CNN-denoised images. To tackle these issues, we characterized CNN output distribution via total uncertainty (i.e., data + model uncertainties) and predictive mean. Local mean-uncertainty-ratio (MUR) was calculated to detect highly unreliable regions in the denoised images. A MUR-driven adaptive local fusion (ALF) process was developed to adaptively merge local predictive means with the original noisy images, thereby improving image robustness. This process was incorporated into a previously validated deep-learning model observer to quantify liver metastasis detectability. For proof-of-concept, the proposed method was established and validated for a ResNet-based CT denoising method. The proposed method consistently improved lesion detectability in challenging conditions such as lower dose, smaller lesion size, or lower contrast. The proposed method has the potential to improve reliability of deep-learning CT denoising and enhance lesion detection.
13405-57
20 February 2025 • 4:40 PM - 5:00 PM PST | Town & Country B
Show Abstract +
Myocardial perfusion imaging (MPI) by single-photon emission computed tomography (SPECT) is a widely used
diagnostic tool for cardiovascular diseases. There is an important need to reduce the length of scanning procedures
and acquisition time to mitigate patient discomfort, motion artifacts, and potential inaccurate diagnoses due to
misalignment between SPECT and computed tomography (CT) scans. Reducing the number of projection views
provides a mechanism to shorten the scanning time. However, reducing the number of projection views in MPI
SPECT introduces noise and artifacts which decreases the diagnostic accuracy. To address this issue, we propose
a detection-task-specific deep-learning method for sparse-view MPI SPECT. The approach incorporates our
understanding of the human visual system within a deep-learning approach to process the sparse-view SPECT
images towards improving performance on defect-detection tasks. We objectively evaluated the proposed method
on the clinical task of detecting cardiac defects in a retrospective study with anonymized clinical data from
patients who underwent MPI studies.
13405-58
20 February 2025 • 5:00 PM - 5:20 PM PST | Town & Country B
Show Abstract +
This study introduces a novel unsupervised deep learning approach NAS-DRP for CT image reconstruction that integrates Neural Architecture Search (NAS) with Deep Radon Prior (DRP). By utilizing a reinforcement learning network and adapting data inconsistency in the Radon domain, NAS-DRP innovatively optimizes network structures for sparse-view CT imaging without relying on paired training data. This methodology addresses the primary challenges of traditional CT imaging, including high radiation doses and the need for extensive data sets, by significantly enhancing image quality through the optimization of upsampling processes. The incorporation of NAS allows for automatic adjustment of network parameters, catering specifically to the unique demands of medical imaging reconstruction. Experimental validations reveal marked improvements in image quality, demonstrating NAS-DRP's potential to transform CT reconstruction practices by reducing operational complexities and costs. This advancement represents a significant stride in medical imaging, offering a scalable solution that could be extended to other imaging modalities.
Program Committee
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences (China)
Program Committee
U.S. Food and Drug Administration (United States), Univ. of Massachusetts Medical School (United States)
Program Committee
GE HealthCare (United States), Lawrence Berkeley National Lab. (United States)
Program Committee
National Institute of Biomedical Imaging and Bioengineering (United States)
Program Committee
Univ. of Wisconsin School of Medicine and Public Health (United States)
Related Courses
SPIE Courses at Medical Imaging
-
See courses on topics related to Physics of Medical Imaging.
-
Pre-registration is required for SPIE Courses.