This PDF file contains the front matter associated with SPIE Proceedings Volume 11513, including the Title Page, Copyright information, Author Index, and Table of Contents.

Anthropomorphic breast phantoms are used to create images that mimic aspects of clinical breast images and are useful in optimization and characterization of breast imaging systems. Here, a full-sized compressed physical breast phantom is designed and manufactured with 100 m resolution, high reproducibility and x-ray properties similar to that of breast tissues. The phantom design is based on a digital model derived from the morphology and distribution of large, medium and small scale fibroglandular and inter-glandular adipose tissue observed in clinical breast computerized tomography (bCT) images. The physical phantom consists of four slabs of a polyamide-12 component that mimics adipose tissue fabricated using selective laser sintering (SLS). The fibroglandular component is a low viscosity resin doped with a small amount of zinc oxide nanoparticles (<110 nm) to increase attenuation. The phantom was imaged on a Senographe Pristina and compared to image simulations of the virtual phantom. The power spectral parameter, β was 3.8±0.2 and 3.9±0.5 for the physical and virtual phantoms in a digital mammogram. The corresponding Laplacian fractional entropy (LFE) averaged 0.22 and 0.14 across the range 0.125–1.29 mm^-1. Very good texture cancellation was obtained in contrast-enhanced spectral mammography.

Purpose: To investigate the possibility of evaluating synthetic mammograms (SM) with a 3D structured phantom combined with model observer scoring. Methods: SM images were acquired on the Siemens Mammomat Revelation in order to set up a human observer study with 6 readers. Regions of interest with lesions (microcalcifications and masses) present and absent were selected for use in a four-alternative forced choice study. Image acquisitions and reading was performed at AEC,½ AEC and 2×AEC dose levels. The percentage correct (PC) results were calculated for all readers together with the standard error of the mean (SEM). A two-layer non-biased Channelized Hotelling Observer (CHO) for lesion detection was used: a two Laguerre-Gauss channel CHO applied first for localization and then an eight Gabor channel CHO for classification. Observer PC results were estimated using a bootstrap method, and the standard deviation (SD) was used as a figure of merit for reproducibility. Results: Following tuning steps, good correlation was found between the MO and human observer results for both microcalcifications and masses, at the three dose levels. The CHO predicted the PC values of the human readers, but with better reproducibility than the human readers. The detection threshold trends of the CHO matched those of the human observers. Conclusion: A two-layer CHO, with appropriate tuning and testing steps, could approximate the human observer detection results for microcalcifications and masses in SM images acquired on a Siemens Revelation DBT systems over three dose levels . The model observer developed is a promising candidate to track imaging performance in SM.

Anthropomorphic digital breast phantoms are an essential part in the development, simulation, and optimisation of x-ray breast imaging systems. They could be used in many applications, such as running virtual clinical trials or developing dosimetry methods. 3D image modalities, such as breast computed tomography (BCT), provide high resolution images to help produce breast models with realistic internal tissue distribution. However, in order to mimic X-ray imaging procedures such as mammography or digital breast tomosynthesis, the breast model needs to be compressed. In this work, we describe a method to generate compressed breast phantoms using a biomechanical finite element (FE) model from BCT volumes, by simulating physically realistic tissue deformation. Unlike prior literature, we propose a new tissue interpolation methodology which avoids interpolating the deformation fields, resulting in the preservation of the breast tissue amount during the compression process and therefore increasing the accuracy of the deformation. In this study, a total of 88 BCT images were compressed in order to obtain a set of realistic phantoms. The information associated with the phantom (i.e. amount of glandular tissue and adipose tissue and total breast volume) is compared before and after compression (showing a correlation R of 0.99). Also, the same metrics were evaluated between compressed phantoms and VolparaTM measurements from breast tomosynthesis images (R=0.81 − 0.85). Furthermore, we include a 3D surface analysis and describe several medical physics applications in which our phantoms have been used: x-ray dosimetry, scattered radiation estimation or glandular tissue assessment.

Purpose: There are currently five FDA approved commercial digital breast tomosynthesis (DBT) systems, all of which have varying geometry and exposure techniques. The aim of this work was to determine if an anthropomorphic breast phantom could be used to systematically compare performance of DBT, full field digital mammography (FFDM) and synthetic mammography (SM) across the systems. Methods: An anthropomorphic breast phantom was created through inkjet printing containing printed masses. The phantom was imaged using automatic exposure control (AEC) settings for that system. Thus, all phantom acquisition settings, and subsequent radiation dose levels, were dictated from the manufacturer settings. A four alternative forced choice reader study was conducted to assess reader performance. Results: Performance in detecting masses was higher with DBT than with FFDM or SM. The difference in proportion correct (PC) was statistically significant for most cases. Additionally, PC of the DBT systems trended with increased gantry span with lowest PC from Hologic and Fuji (both 15°), then both GE systems (25°), and highest for Siemens (50°). Conclusions: A phantom containing masses was imaged on five commercially available DBT systems across 3 states. A 4AFC study was performed to assess performance with FFDM, DBT, and SM across all systems. Overall detection was highest using DBT, with improvement as the gantry span increased. This study is the first of its kind to use an inkjet based physical anthropomorphic phantom to assess performance of all five commercially available breast imaging systems.

Patient-specific finite element (FE) models of the breast have received increasing attention due to the potential capability of fusing information from different image modalities. During the Magnetic Resonance Imaging (MRI) to X-ray mammography (MG) registration procedure, a FE model is compressed mimicking the mammographic acquisition. To develop an accurate model of the breast, the elastic properties and stress-strain relationship of breast tissues need to be properly defined. Several studies (in vivo and ex vivo experiments) have proposed a range of values associated to the mechanical properties of different tissues. This work analyse the elastic parameters (Young Modulus and Poisson ratio) obtained during the process of registering MRI to X-ray MG images. Position, orientation, elastic parameters and amount of compression are optimised using a simulated annealing algorithm, until the biomechanical model reaches a suitable position with respect to the corresponding mammogram. FE models obtained from 29 patients, 46 MRI-MG studies, were used to extract the optimal elastic parameters for breast compression. The optimal Young modulus obtained in the entire dataset correspond to 4.46 ± 1.81 kP a for adipose and 16.32 ± 8.36 kP a for glandular tissue, while the average Poisson ratio was 0.0492 ± 0.004. Furthermore, we did not find a correlation between the elastic parameters and other patient-specific factors such as breast density or patient age.

Some image processing techniques for digital mammography (DM) and digital breast tomosynthesis (DBT) require or assume prior knowledge of the three-dimensional (3D) shape of the compressed breast. Our goal is to characterize the breast shape curvature during mechanical compression for DM and DBT acquisition, and to study its dependencies on relevant patient features to improve this prior knowledge. For this, patients were recruited to undergo 3D breast surface scanning during breast compression for their clinical DBT examination. This surface scanning system included two sets of a digital projector with a pair of stereoscopic structured light cameras, positioned at opposite sides of the DBT system. Features from the patients were extracted from the DBT images while the 3D surfaces were used as input to a previously developed principal component analysis (PCA) breast shape model. From the cases included in this study, 70 scans with full breast coverage and without artefacts were selected for this interim analysis. The enhanced coverage with the stereoscopic setup resulted in improved characterization of the breast shape curvature, yielding an asymmetry in the curvature between medial and lateral side. Linear correlation between the first PCA component and the thickness of the breast was found, but not for the other components. A multiple linear regression analysis was applied, finding no significant correlation between curve shape and patient characteristics. Searching for other factors that could be used to predict the breast shape when compressed, and testing for non-linear correlations will be addressed as a next step.

Purpose: To investigate the potential of uncertainty analysis asthe first step to explore the relation between lesions texture, in single energy temporal contrast-enhanced mammography (SET), and immunohistochemistry (IHC) status of breast cancer. Methods: Texture features (TF) extracted from the co-occurrence matrix were considered. We studied three sources of uncertainty: stability of the mammography unit, misalignment between pre- and post-contrast images, and manual delineation of suspicious regions. The first two sources were analyzed using phantoms. For uncertainty due to manual delineation, three different radiologists segmented 33 malignant lesions on SET studies. Two segmentation criteria were evaluated: to draw around the lesion border, and to select a focal region with the greatest suspicion of malignancy. Inter- and intra-observer agreement were evaluated in terms of the intra-class correlation coefficient (ICC) and the Pearson correlation coefficient (PCC). The relation between texture features and IHC status was explored. Results: Misalignment was the major source of uncertainty, followed by lesion delineation and the stability of the mammography unit. There was good inter-observer (ICC>0.7) and intra-observer (PCC>0.8) agreement among TF obtained from regions around the lesion border; however, TF from focal regions only agreed in terms of mean value and correlation. Texture analysis predicted the presence of hormone receptors and a high proliferation rate moderately better than an educated. The texture features that conducted to the best prediction models were the mean value, Imc2 and average contrast. Conclusions: Uncertainty evaluation improves textures analysis and assessment of the prediction model. A wider range of imaging features could improve the prediction of (IHC) status.

In this paper, radiomic features are used to validate the textural realism of two anthropomorphic phantoms for digital mammography. One phantom was based off a computational breast model; it was 3D printed by CIRS (Computerized Imaging Reference Systems, Inc., Norfolk, VA) under license from the University of Pennsylvania. We investigate how the textural realism of this phantom compares against a phantom derived from an actual patient’s mammogram (“Rachel”, Gammex 169, Madison, WI). Images of each phantom were acquired at three kV in 1 kV increments using auto-time technique settings. Acquisitions at each technique setting were repeated twice, resulting in six images per phantom. In the raw (“FOR PROCESSING”) images, 341 features were calculated; i.e., gray-level histogram, co-occurrence, run length, fractal dimension, Gabor Wavelet, local binary pattern, Laws, and co-occurrence Laws features. Features were also calculated in a negative screening population. For each feature, the middle 95% of the clinical distribution was used to evaluate the textural realism of each phantom. A feature was considered realistic if all six measurements in the phantom were within the middle 95% of the clinical distribution. Otherwise, a feature was considered unrealistic. More features were actually found to be realistic by this definition in the CIRS phantom (305 out of 341 features or 89.44%) than in the phantom derived from a specific patient’s

Computer aided diagnosis (CADx) systems for digital mammography mostly rely on 2D techniques. With the emergence of three-dimensional (3D) breast imaging modalities, such as digital breast tomosynthesis (DBT), there is an opportunity to create 3D models and analyze 3D features to classify microcalci€cations (MC) clusters to help the early detection of breast cancer. We adopted the 3L algorithm for implicit B-spline (IBS) €‹ing to investigate the robustness of 3D models of microcalci€cation (MC) clusters for classifying benign and malignant cases. Point clouds were initially generated from tomosynthesis slices. Two additional o‚set points were generated to support the original point clouds for detailed 3D modeling. Before €‹ing the splines, the point clouds were normalized into a unit cube la‹ice. A‰er modeling individual MCs into a unit cubic la‹ice, they are all located in a 3D space according to their spatial location in the tomosynthesis images to form a cluster. Features were extracted from the 3D model of MC clusters. With selected features we obtained 80% classi€cation accuracy.

Medical image synthesis has gained a great focus recently, especially after the introduction of Generative Adversarial Networks (GANs). GANs have been used widely to provide anatomically-plausible and diverse samples for augmentation and other applications, including segmentation and super resolution. In our previous work, Deep Convolutional GANs were used to generate synthetic mammogram lesions, masses mainly, that could enhance the classification performance in imbalanced datasets. In this new work, a deeper investigation was carried out to explore other aspects of the generated images evaluation, i.e., realism, feature space distribution, and observer studies. t-Stochastic Neighbor Embedding (t-SNE) was used to reduce the dimensionality of real and fake images to enable 2D visualisations. Additionally, two expert radiologists performed a realism-evaluation study. Visualisations showed that the generated images have a similar feature distribution of the real ones, avoiding outliers. Moreover, the Receiver Operating Characteristic (ROC) study showed that the radiologists could not, in many cases, distinguish between synthetic and real lesions, giving accuracies between 51% and 59% using a balanced sample set.

Breast cancer screening is predominantly performed using digital mammography (DM), but higher sensitivity has been demonstrated with digital breast tomosynthesis (DBT). A partial DBT screening in selected groups with a clear benefit from DBT might be more feasible than a full implementation, and using artificial intelligence (AI) to select women for DBT might be a possibility. This study used data from Malmö Breast Tomosynthesis Screening Trial, where all women prospectively were examined with separately read DM and DBT. We retrospectively analysed DM examinations (n=14768) with a breast cancer detection software and used the provided risk score (1-10) for risk stratification. We tested how different score thresholds for adding DBT to an initial DM affects the number of detected cancers, additional DBT examinations needed, detection rate, and false positives. If using a threshold of 9.0, 25 (26 %) more cancers would be detected compared to using DM alone. Of the 41 cancers only detected on DBT, 61 % would be detected, with only 1797 (12 %) of the women examined with both DM and DBT. The detection rate for the added DBT would be 14/1000 women, while the false positive recalls would be increased with 58 (21 %). Using DBT only for selected high gain cases could be an alternative to a complete DBT screening. AI could be used for analysing DM to identify high gain cases, where DBT can be added during the same visit. There might be logistical challenges and further studies in a prospective setting are necessary.

We investigated whether a deep learning-based artificial intelligence (AI) system can be used to improve breast cancer screening workflow efficiency by making a pre-selection of likely normal screening mammograms where double-reading could be safely replaced with single-reading. We collected 18,015 consecutively acquired screening exams, the independent reading assessments by each radiologist of the double reading process, and the information about whether the case was recalled and if so the recall outcome. The AI system assigned a 1-10 score to each screening exam denoting the likelihood of cancer. We simulated the impact on recall rate, cancer detection rate, and workload if single-reading would have been performed for the mammograms with the lowest AI scores. After evaluating all possible AI score thresholds, it was found that when AI scores 1 to 7 are single read instead of double read, the cancer detection rate would have remained the same (no screen-detected cancers missed –the AI score is low but the single-reader would recall the exam), recall rate would have decreased by 11.8% (from 5.35% to 4.79%), and screen reading workload would have decreased by 32.6%. In conclusion, using an AI system could improve breast cancer screening efficiency by pre-selecting likely normal exams where double-reading might not be needed.

Aim: To investigate patterns of false-positive decisions in mammography. The relationship between frequently falsely recalled mammograms, breast density, assigned confidence score of malignancy and presenting mammographic features of the case were explored. Method: There were 2 phases to this experiment. In Phase 1, 200 cases (180 normal, 20 cancer containing) were read by 5 Australian radiologists under three recall conditions: free recall, 15% recall rate and 10% recall rate. The readers assigned any suspicious cases a confidence score of malignancy. The most frequently recalled normal cases were identified via pattern analysis and assess for chance finding. These cases underwent case feature analysis including mammographic breast density and assigned confidence score of malignancy. In Phase 2: Seven expert breast radiologists were asked to review the cases knowing they were normal and rate rating each case for displaying “normal” and/or “benign” mammographic features and give a difficulty rating. Results: 18 normal cases were identified as repeatedly recalled across all recall rates. There was no correlation between mammographic breast density and FP frequency. Under the 10% recall rate condition, normal cases in the extremely difficult category were associated with significantly high confidence scores of malignancy (X2(2) =6.701, p=0.035). There was moderate correlation between the cohorts on case difficulty and the survey results indicated that benign disease was often marked as malignant. Conclusion: There was a clear pattern of normal mammograms persistently recalled. FP cases were not associated with higher mammographic breast density but did receive high confidence scores of malignancy. Radiologists agreed that highly difficult cases did not fit with their knowledge of normal variants. The phenomenon of repetitive occurrences of FP experience for screened women warrants further research on the analysis of the textural characteristics of normal mammographic cases.

2427 women (mean age 60.0 years; range 20-94 years) were screened with DBT and ABVS from May 2014 to December 2018. 5375 women (mean age 57.3 years; range 40-84 years) were screened with MMG and HUS from February 2004 to December 2013. Number of recall cases (rate), unidentified cases, further exams, identified outcome cases, breast cancer cases, false positive cases, positive predictive value (PPV), cancer detection rate, implementation rate of further exams (fine needle aspiration (FNA), core needle biopsy (CNB), vacuum-assisted breast biopsy(VAB)) were evaluated between two groups. The recall rate was lower in women with DBT and ABVS in comparison with women with MMG and HUS (7.9 % vs 8.2%; p>0.05). Detection rate of breast cancers was higher in women with DBT and ABVS than women with MMG and HUS (1.6% vs 1.2%; p>0.05). Non-cancer rate among women receiving invasive procedures was lower in women with DBT and ABVS in comparison with women with MMG and HUS (0.9% vs 1.0%; p>0.05). PPV was significantly higher in women with DBT and ABVS in comparison with women with MMG and HUS (19.9% vs 14.5%; p=0.045). DBT and ABVS screening improved the cancer detection rate with significant higher PPV in comparison with MMG and HUS for breast cancer screening.

This work compares estimates of the radiation dose in mammography obtained using three different fibroglandular tissue distributions. Ninety volumetric images of patient breasts were acquired with a dedicated breast CT system and the voxels automatically classified as containing skin, adipose, or glandular tissue. The classified images underwent simulated mechanical compression to mimic the mammographic cranio-caudal acquisition. The voxels containing fibroglandular and adipose tissue were then distributed in the breast phantoms following three different methods: patient-based (i.e., maintaining the original distribution), homogeneous (i.e., each voxel is a homogeneous mixture of adipose and glandular tissue) and newly-proposed continuous (i.e., the glandular tissue is distributed according to a general model, derived from the patient breast CT data). All breast phantoms were used in Monte Carlo simulations to estimate the radiation dose. The results show that the doses estimated using the continuous fibroglandular tissue distribution agree within 3% of the doses estimated using the heterogeneous patient-based distribution, and that it leads to a dose reduction of 27% compared to the homogeneous distribution.

Stereotactic breast biopsy (SBB) is a common clinical procedure for suspicious breast lesion analysis. With the arrival of DBT-guided biopsy systems, the clinical performance of such procedures has improved enormously since breast lesions are better detected. However, little information is found in the literature regarding the patient’s radiation dose during these clinical procedures. This work presents, for the first time, a first approach to estimate the mean glandular dose (MGD) within the biopsy window for 101 patients who underwent breast biopsy in a commercially available DBT-guided prone table. This study is supported by the calculation of normalised glandular dose (DgN) coefficients from Monte Carlo simulations. Preliminary results show that the total MGD of the biopsy procedure varies between 10.2 mGy and 19.2 mGy for patients with breast thickness between 2 cm and 8 cm. Furthermore, a great variability in the number of acquisitions (tomo scan or stereo projections) of the biopsy procedure was observed. For the investigated system, MGD for DBT-guided breast biopsies are, for 5-6 cm thick breasts, around 23% lower than MGD observed in stereo biopsy procedures. The proposed method represents a first approach towards a full dose estimation of DBTguided breast biopsy procedures.

Purpose: To test the association between contrast-to-noise ratio (CNR) measurements made on digital mammograms (DM), human reader performance in a lesion detection task using the same images, and image quality (IQ) as predicted by phantom measurements. Methods: DM from 162 women were evaluated for their CNR using a novel metric for application on clinical images. The original unprocessed images were tested (100% dose), as well as the same images after processing to simulate a 50% and 25% relative dose level. IQ measurements from a CDMAM phantom images, as well as human reader calcification cluster detectability ratings on the clinical image set for the three treatments were used to provide ground truth for human lesion detection performance. Analysis was performed to test for association between DM image CNR at the three dose levels, the CDMAM measurements, and reader performance as quantified by a reader-averaged jack-knifed free response operating characteristic (JAFROC) figure of merit (FoM). Results: The clinical image CNR was strongly correlated with the JAFROC FoM and CDMAM threshold gold thicknesses (r=0.98, and r=0.99 @ 0.25 mm, r=0.94 @ 0.1 mm discs, respectively). On a per-image basis, strong associations between CNR and measures of beam quality and exposure were also found that indicate sensitivity to imaging technique factors while remaining independent of signal variations due to breast parenchyma. Conclusions: Using a clinical image CNR it is possible to objectively predict IQ in mammographic images. As such, this metric could provide a means to perform a practical continuous DM system performance monitoring.

The image quality of contrast-enhanced digital breast tomosynthesis (CEDBT) is degraded by scatter radiation. Scatter correction can improve the object contrast and reduce the cupping artifacts, but the image quality is limited by the increased image noise. In this study we investigate the effect of scatter correction on image noise in CEDBT. A scatter correction method based on image convolution with scatter-to-primary ratio kernel was applied. We analyzed the noise power spectrum (NPS) for CEDBT projection images before and after scatter correction using CIRS breast phantoms and evaluated the signal-difference-to-noise ratio (SDNR) of the iodine objects after image reconstruction. We applied image filtering to reduce image noise after scatter correction for phantom and clinical images. A deep learning based denoising technique was applied to further reduce the image noise for clinical images. Our results show that the scatter correction increases the image noise in dual-energy subtracted images, and the improvement in SDNR from scatter correction is limited. Noise reduction applied after scatter removal can regain the benefit in SDNR from scatter correction and further improve the visualization of contrast enhancement in CEDBT.

Noise negatively impacts the detection and characterization of lesions in mammography. While denoising filters may be used to suppress noise, they might also negatively affect the conspicuity of small lesions due to signal blurring and smearing. In previous works, we designed and validated a denoising pipeline, dedicated to mammography, capable of suppressing noise and avoiding excessive blur and smear. This is achieved by a fine-tuned noisy-denoised image blending step, which leverages a Poisson-Gaussian noise model. In the current work, we investigate the impact of the denoising pipeline on the localization of low contrast microcalcification clusters. To this end, a human observers study was conducted with a team of five medical physicists with experience in breast imaging. First, in the pilot study, we defined the limit of contrast for the localization task with and without the application of the denoising pipeline. Next, we investigated the effect of the denoising on the localization of microcalcification clusters. Clinical patient cases with dense breasts and simulated microcalcification clusters were used throughout this study to emulate challenging cases and to guarantee fine control over the lesion’s contrast. The results from six readers show that the limit of localization occurred at the contrasts 0.090 and 0.079 without and with denoising, respectively. The average correct localization rate was 77% and 81% without and with denoising, respectively. Thus, the results show that the readers were able to correctly locate significantly less conspicuous lesions (p<0.05), and also performed significantly better localizing microcalcification clusters (p<0.05) when the denoising pipeline was applied.

This study explored whether having a better performance in usual presentation condition, more years of experience, and higher volume of annual mammogram assessment make a radiologist better at perceiving the gist of the abnormal on a mammogram. Nineteen radiologists were recruited for two experiments. In the first one (gist experiment), the initial impressions of the radiologists were collected based on a half-second image presentation on a scale of 0 (confident normal) and 100 (confident abnormal). In the second one, radiologists viewed similar set of cases using BreastScreen Reader Assessment Strategy platform and rated each case on a scale of 1-5. Using Spearman correlation, we explored if the area under receiver operating characteristics curve (AUC) in two experiments were correlated. Radiologists were also grouped based on variables describing their experience levels and workload and their performance in both experiments were compared among the groups. The AUC values in the gist experiment was not significantly correlated to the AUC values in the normal reporting experiment (Spearman correlation=0.183, p-value=0.453). Radiologists’ performances under the normal reporting conditions, was linked to the number of cases per week (p=0.044), number of hours per week currently spent reading mammograms(p=0.028), and number of years they have been reading mammograms (p=0.041). However, none of the variables reached a p-value<0.05 for the AUC of the gist experiment. The results suggest that further studies should be done to establish relationships between the gist response and radiologists’ characteristics since being a high-performing radiologist, highly experienced radiologist, or reading high volume of mammograms does not indicate superior capability when perceiving the gist of the abnormal.

Our purpose was to determine if the lipid, water, and protein lesion composition (3CB), combined with computer-aided detection (CAD) had higher biopsy malignancy specificity than CAD alone. High and low-kVp full-field digital 3CB mammograms were acquired on women with suspicious mammographic lesions (BIRADS 4) and that were to undergo biopsy. Radiologists delineated 673 lesions (98 invasive ductal cancers (IDC), 60 ductal carcinomas in situ (DCIS), 103 fibroadenomata (FA), and 412 benign (BN)) on the diagnostic mammograms using the pathology report to confirm location. The diagnostic mammograms were processed by iCAD SecondLook software using its most sensitive setting to create to further delineations and probabilities of malignancy. The iCAD delineated a total of 375 annotation agreeing regions that were classified as either masses or calcification cluster. The 3CB algorithm produced lipid, water, and protein thickness maps for all ROIs and peripheral rings from which 84 compositional input features were derived. A neural network (3CBNN) was trained with cross-validation on 80% of the data to predict the lesion type. Biopsy pathology served as the gold standard outcome. IDC and DCIS predicted probabilities were summed together to obtain a probability of malignancy which was evaluated against the iCAD probabilities using the area under the ROC curves. On a holdout test set, 20% of the data, the iCAD's output alone had an AUC of 0.61 while the 3CBNN’s AUC was 0.73. We conclude that compositional information provided by the 3CB algorithm contains important diagnostic information that can increase specificity of CAD software.

Deep learning models have reached superior results in various fields of application, but in many cases at a high cost of processing or large amount of data available. In most of them, specially in the medical field, the scarcity of training data limits the performance of these models. Among the strategies to overcome the lack of data, there is data augmentation, transfer learning and fine-tuning. In this work we compared different approaches to train deep convolutional neural network (CNN) to automatically detect architectural distortion (AD) in digital mammography. Although several computer vision based algorithms were designed to detect lesions in digital mammography, most of them perform poorly while detecting AD. We used the VGG-16 network pre-trained on ImageNet database with progressive fine-tuning to evaluate its performance on AD detection over a database of 280 images of clinical mammograms. Finally, we compared the results with a custom CNN architecture trained from scratch for the same task. Results indicated that a network with transfer learning and certain level of fine-tuning reaches the best results for the task (AUC = 0.89) compared with the other approaches, but no statistically significant difference was found between the best results using different amount of data augmentation and also compared to the custom CNN.

In this study we used a large previously built database of 2,892 mammograms and 31,650 single mammogram radiologists’ assessments to simulate the impact of replacing one radiologist by an AI system in a double reading setting. The double human reading scenario and the double hybrid reading scenario (second reader replaced by an AI system) were simulated via bootstrapping using different combinations of mammograms and radiologists from the database. The main outcomes of each scenario were sensitivity, specificity and workload (number of necessary readings). The results showed that when using AI as a second reader, workload can be reduced by 44%, sensitivity remains similar (difference -0.1%; 95% CI = - 4.1%, 3.9%), and specificity increases by 5.3% (P<0.001). Our results suggest that using AI as a second reader in a double reading setting as in screening programs could be a strategy to reduce workload and false positive recalls without affecting sensitivity.

Breast Magnetic Resonance Imaging (MRI) is recognized as the most sensitive imaging method for the early detection of breast cancer in women who carry a lifetime risk for breast cancer higher than or equal to 20%. Given the aggressive biology of cancers in this population, early detection is crucial for a favorable prognosis. This study aimed to use artificial intelligence for the detection of lesions at the earliest stage in high-risk women. A Generative Adversarial Network (GAN) detected lesions in breast MR data by quantifying anomaly as divergence from healthy breast tissue appearance. First, follow-up images of patients were aligned and the breast was segmented automatically. Then, the GAN created a model of healthy variability of appearance change during follow-up in 64x64-sized image patches sampled only at healthy tissue locations in follow-up image sequences. During the assessment of new data, each image position was compared with the model yielding an anomaly score. On a image patch level, we evaluated if this anomaly score identifies confirmed lesions, as well as lesionfree regions, where lesions appear during later follow-up studies. In the first experiment of lesion detection, a mean sensitivity of 99.5% and a mean specificity of 84% was achieved. When applying the model to studies denoted as lesion-free, subsequently occurring lesions were predicted with a mean sensitivity of 92.7% and a mean specificity of 78.8%.

Dual energy contrast-enhanced digital breast tomosynthesis (CEDBT) uses weighted subtraction of two energy spectra to highlight tumor angiogenesis with uptake of iodinated contrast agent. The high energy scan contains more severe scatter radiation than regular low energy DBT. The purpose of this study is to develop a convolutional neural network (CNN) based scatter correction method for dual energy CEDBT in both craniocaudal (CC) view and mediolateral oblique (MLO) view. Anthropomorphic digital breast phantoms with various glandularity and 3D shape were generated using the VICTRE software tool developed by the FDA. The pectoralis muscle layer was inserted into the phantoms for MLO view. Projection images with and without scatter radiation were simulated using Monte Carlo (MC) simulation code of VICTRE, meeting the prototype Siemens Mammomat Inspiration CEDBT system with 300 μm thick a-Se detector, 25 projections within 46-degree angular range. Scatter radiation ground truth was generated from MC simulated projection images to train CNN. Two separate U-net CNNs were trained to predict scatter radiation maps. Mean absolute percentage error (MAPE) was used as the loss function. The average MAPE of this method is less than 3 % from the ground truth of MC simulation. The proposed scatter correction method was then applied to clinical cases, demonstrating the reduction of cupping artifact and the improvement in contrast object conspicuity.

Specimen x-ray imaging provides important information on the margin of surgically excised tissue as well as radiologic and pathologic correlation of the lesion. Similar to breast imaging, where mammograms are digitally processed to enhance readability and lesion conspicuity, specimen images are also processed and enhanced. However, specimen image processing is made challenging by the diversity of specimen containers that are commercially available, compounded by variations in specimen size. In this work, we demonstrate our specimen container and size classification system based on a simple convolutional neural network (CNN), trained to identify the container type. This system allows for automated image processing of the supported container types. A dataset consisting of 1428 HIPAA and IRB-complaint anonymized specimen images were collected. We prepared a simple CNN for image classification with 3 convolutional and 3 fully connected layers, and evaluated the performance based on three comparison metrics. Each network was analyzed in terms of accuracy, multi-class AUC, and via a confusion matrix. The best performing classifier, determined via cross validation, was then used for testing, and evaluated with the same three metrics. The results of training and tuning within cross validation showed that the specimen classes are easily differentiable with this simple convolutional neural network structure. During testing, the network was able to achieve an accuracy of 95.8±4.0%, and an AUC of 0.9763±0.0001.

A Phase-Contrast breast CT facility based on a high-resolution CdTe photon-counting detector is under development at Elettra, the Italian Synchrotron Radiation (SR) facility in Trieste. The CT system exploits propagation-based phasecontrast imaging and phase-retrieval algorithm. The voxel size is 57×57×50 μm³ and the delivered MGDs, about 5 mGy, are comparable with clinical breast CT systems. In the present contribution, the comparisons between histological breast cancers and full breast CT images are presented from samples of breast mastectomy. The high resolution of the breast CT images and low noise due to the phase contrast allow a very fine matching between x-ray CT and histology at acceptable delivered doses.

Propagation-based phase-contrast CT (PB-CT) is a novel imaging technique that visualises variations in both X-ray attenuation and refraction. This study aimed to compare the clinical image quality of breast PB-CT using synchrotron radiation with conventional absorption-based CT (AB-CT), at the same radiation dose. Seven breast mastectomy specimens were scanned and evaluated by a group of 14 radiologists and medical imaging experts who assessed the images based on seven radiological image quality criteria. Visual grading characteristics (VGC) were used to analyse the results and the area under the VGC curve was obtained to measure the differences between the two techniques. For six image quality criteria (overall quality, perceptible contrast, lesion sharpness, normal tissue interfaces, calcification visibility and image noise), PB-CT images were superior to AB-CT images of the same dose (AUC_VGC: 0.704 to 0.914, P≤.05). For the seventh criteria (artefacts), PB-CT images were also rated better than AB-CT images (AUC_VGC: 0.647) but the difference was not significant. The results of this study provide a solid basis for future experimental and clinical protocols of breast PB-CT.

We are developing a prototype system for simultaneous digital breast tomosynthesis (DBT) and mechanical imaging (MI). MI maps the local pressure distribution during clinical exams, to distinguish breast abnormalities from the normal tissue. Both DBT alone, and MI when combined with digital mammography, have demonstrated the ability to reduce false positives; however, the benefit of combining DBT with MI has not been investigated. A practical limitation in simultaneous DBT and MI is the presence of the MI sensor in DBT images. Metallic elements of the sensor generate noticeable artifacts, which may interfere with clinical analysis. Previously, we shown that the sensor artifacts can be reduced by flat fielding, which combines projections of the sensor acquired with and without the breast. In this paper we evaluate the flat fielding by assessing artifact reduction and visibility of breast abnormalities. Images of a physical anthropomorphic breast phantom were acquired using a clinical wide-angle DBT system. Visual evaluation was performed by experienced medical physicists. Image quality descriptors were calculated in images with and without flat fielding. To evaluate the visibility of abnormalities we estimated the full width at half maximum (FWHM) for calcifications modeled in the phantom. Our preliminary results suggest a substantial reduction of artifacts by flat fielding (on average 83%). Few noticeable artifacts remain near the breast edge, in the reconstructed image with the sensor in focus. We observed a 17% reduction in the FWHM. Future work would include a detailed assessment, and method optimization using virtual trials as a design aid.

Our previous work explored the use of super-resolution as a way to improve the visibility of calcifications in digital breast tomosynthesis. This paper demonstrates that there are anisotropies in super-resolution throughout the reconstruction, and investigates new motion paths for the x-ray tube to suppress these anisotropies. We used a theoretical model of a sinusoidal test object to demonstrate the existence of the anisotropies. In addition, high-frequency test objects were simulated with virtual clinical trial (VCT) software developed for breast imaging. The simulated objects include a lead bar pattern phantom as well as punctate calcifications in a breast-like background. In a conventional acquisition geometry in which the source motion is directed laterally, we found that super-resolution is not achievable if the frequency is oriented in the perpendicular direction (posteroanteriorly). Also, there are positions, corresponding to various slices above the breast support, at which super-resolution is inherently not achievable. The existence of these anisotropies was validated with VCT simulations. At locations predicted by theoretical modeling, the bar pattern phantom showed aliasing, and the spacing between individual calcifications was not properly resolved. To show that super-resolution can be optimized by re-designing the acquisition geometry, we applied our theoretical model to the analysis of new motion paths for the x-ray tube; specifically, motions with more degrees of freedom and with more rapid pulsing (submillimeter spacing) between source positions. These two strategies can be used in combination to suppress the anisotropies in super-resolution.

The value of adding ^99mTc- sestamibi MBT to the current clinical standard of DBT plus 2D digital mammography (MM) was assessed. Images were acquired using a dual modality tomosynthesis (DMT) scanner designed to obtain superimposable DBT and MBT images. Seventy-five subjects with 83 biopsied lesions were scanned prior to biopsy. A blinded MQSA-certified breast radiologist with limited nuclear medicine (NM) experience viewed the images in the following sequence: 1) DBT alone, 2) add MM, and 3) add MBT (equivalent to DMT+MM). MM images were from each subject’s most recent clinical mammographic exam. At each stage, all findings were scored using a 5-point suspicion scale ranging from 1=definitely benign, to 5=definitely malignant. Independently, a blinded, experienced NM radiologist scored all MBT scans without access to the DBT or MM images, using the same suspicion scale. The NM results were provided to the breast radiologist reader following their 3-stage evaluation, and a fourth suspicion score was recorded for all findings. Using location-confirmed biopsy results as ground truth, ROC curves and the areas under the curves, A_z were generated for each of the four stages, and for MBT alone. Compared to DBT+MM, the changes in Az for MBT alone, DBT, and DMT+MM were +21.4% (p<0.02), -22.1% (p<0.01), +25.2% (p<0.002), respectively. Addition of the NM report to DMT+MM had no measurable effect on ROC shape or A_z value. These results suggest that hybrid tomosynthesis can potentially improve DBT diagnostic performance; that breast radiologists with limited nuclear medicine experience might nevertheless effectively utilize MBT information; and that stand-alone MBT could be a valuable complementary tomographic modality.

Machine learning (ML) has made great advancements in imaging for breast cancer detection, including reducing radiologists read times, yet its performance is still reported to be at best similar to that of expert radiologists. This leaves a performance gap between what is desired by radiologists and what can actually be achieved in terms of early detection, reduction of excessive false positives and minimization of unnecessary biopsies. We have seen a similar situation with military intelligence that is expressed by operators as “drowning in data and starving for information”. We invented Upstream Data Fusion (UDF) to help fill the gap. ML is used to produce candidate detections for individual sensing modalities with high detection rates and high false positive rates. Data fusion is used to combine modalities and dramatically diminish false positives. Upstream data, that is closer to raw data, is hard for operators to visualize. Yet it is used for fusion to recover information that would otherwise be lost by the processing to make it visually acceptable to humans. Our research with breast cancer detection involving the fusion of Digital Breast Tomosynthesis (DBT) with Magnetic Resonance Imaging (MRI) and also the fusion of DBT with ultrasound (US) data has yielded preliminary results which lead us to conclude that UDF can help to both fill the performance gap and reduce radiologist read time. Our findings suggest that UDF, combined with ML techniques, can result in paradigm changes in the achievable accuracy and efficiency of early breast cancer detection.

Computer-Aided Detection (CAD) systems are used to help readers in interpreting screening mammograms. Traditional use of CAD in mammography involves an expert reader searching the image initially unaided, and then once again with the aid of CAD prompts that automatically indicate suspicious regions. An alternative approach is interactive CAD, where prompts are only displayed when readers query a suspicious region for which a prompt is available. These prompts are typically displayed with a given confidence of malignancy for that region. Two non-expert observer studies of interactive CAD were conducted to investigate its effect on the visual search of synthetic images containing microcalcification clusters. Experiment 1 (n=44) used no-CAD and interactive CAD conditions, whereas Experiment 2 (n=43) used interactive CAD in both conditions but in one there was an additional ‘image score’ denoting the likelihood that an image contained a cluster. In both experiments, the addition of interactive CAD (Experiment 1) and an image score (Experiment 2) did not change sensitivity or specificity compared to no-CAD and interactive CAD-alone, respectively. In Experiment 1, the higher the confidence value for a given prompt, the more likely a participant was to act on it. This effect was stronger for true prompts than false prompts. In Experiment 2, participants spent longer viewing images with higher image scores. When no prompt was available, they were more likely to make false positive errors on those images. However, decisions made on available prompts were influenced primarily by confidence values of the prompts rather than overall image score.

Considering the rapid rise in breast cancer incidence in China and lack of calibrated breast cancer prediction models for the Chinese female population, developing a breast cancer model targeting the Chinese women is necessary. This study aimed at generating a breast cancer risk prediction model for Chinese women. A total of 1079 (85 images contralateral to a cancer and 994 cases without breast cancer) women were recruited from Fudan University Shanghai Cancer Centre. For each case, we collected sixteen demographic variables such as age, BMI, number of children, family history of breast cancer, and age at menarche. Moreover, the dense tissue was automatically segmented by AutoDensity. A set of quantitative features were extracted from the dense area. Using the 80th percentile of intensity values in the dense area, the segmented area was thresholded again and the second set of computer-extracted features was calculated. The features, i.e. the demographic variables, and texture features extracted from the mammographically dense areas of the image, have been fed into an ensemble of 250 decision trees, whose results were combined using RUSBoost. The classifier achieved an AUC of 0.88 (CI: 0.84 - 0.91) for identifying high-risk images. Therefore, adopting such model might lead to the augmentation of discriminatory power of currently-used risk prediction models. However, it should be noted that the cancer cases were retrieved from the diagnostic environment (not screening) and further validation on a dataset from a screening set-up will be required.

Breast artery calcification (BAC) is increasingly recognized as a specific marker of medial calcification and may help to identify risk factors of medial artery calcification. Amongst these are high age, diabetes mellitus, hypertension and chronic kidney disease (CKD). Present retrospective observational cohort study focused on the latter patient group with CKD and aimed to define the prevalence and progression rate of BAC in chronic kidney disease (CKD) patients across stages of disease, to define clinical and biochemical correlates of BAC and to explore the association of BAC with incident cardiovascular morbidity and mortality. The main findings of the present observational study are as follows: (a) BAC is common in CKD and its prevalence, severity and rate of progression increase parallel to the degree of kidney dysfunction; (b) inflammation and hyperphosphatemia are (nontraditional) risk factors for BAC in CKD patients; and (c) BAC associates with a dismal cardiovascular outcome in renal transplant recipients. In conclusion, BAC is common among CKD patients, progresses at a slower pace in Tx patients as compared to CKD5D patients, and associates with dismal cardiovascular outcomes. BAC score, kidney function and serum phosphate at baseline seem to be important determinants of progression. BAC is not routinely mentioned in mammogram reports, while the measurement of BAC may offer a personalized, non-invasive approach to risk-stratify CKD patients for cardiovascular disease at no additional cost or radiation since a majority of women over the age of 40 undergo regular breast cancer screening.

For patients with early-stage breast cancer, the axillary lymph node (ALN) metastasis status is one of the important indicators in breast cancer staging and prognosis. In this study, a computer-aided prediction (CAP) system based on the ultrasound image using the deep learning method to determine the ALN status in breast cancer. In this study, the US imaging database contained 153 malignant tumor images which confirmed by histologically examine, and either SNB or ALND confirmed the axillary metastasis status. The Mask R-CNN method is used to indicate the tumor location and extract the tumor region. After the tumor region segmentation process, we obtained the surrounding tissue region (1 mm and 2 mm), which might include implicit information of the tumor metastasis mechanism. Finally, the convolution neural network (CNN)-based classifier is used to predict the ALN metastasis status using segmented images. In the experiments, the results show that the combined region (tumor with 2 mm surrounding tissue) image has the highest predict performance. The accuracy, sensitivity, specificity, and the area index (Az value) under the receiver operating characteristic (ROC) curve for the CAP system were 77.12%, 66.10%, 84.04%, and 0.7592 for using combined region (tumor with 2 mm surrounding tissue) images. These results indicated that the proposed CAP system can be helpful to determine the ALN status in patients with early-stage breast cancer.

A complete software platform based on anthropomorphic breast models used with both planar and three-dimensional phase contrast breast imaging is presented and subjectively validated. For the development of the platform, tests with three anthropomorphic breast phantoms, available both in computational and physical form, were designed and implemented. The models are characterized with different complexity: two phantoms are with spheres and one anthropomorphic. Further on, two of the physical breast models were created with the use of 3D printing techniques. These phantoms with thickness of 40 mm and 31 mm, respectively, were based on digital phantoms created with in-house developed software tools. The third physical breast phantom is the L1 phantom developed at Katholieke Universiteit Leuven with 58 mm thickness. Based on this physical phantom, a computational one was created. The three physical breast phantoms were imaged at ID17 biomedical imaging line at ESRF. Two acquisition setups were used: planar and limited angle tomography modes. Simulated and experimental planar and three-dimensional images were compared in terms of visual reproducibility. Results showed that phantoms characterized with more simple structure produce subjectively similar experimental and simulation appearance in terms of object reproduction and similar edge effects. The thicker phantom demonstrated lower visual coincidence between the two types of planar images, due to higher thickness and higher energy incident beam. The results of this study will be used in the design of new experimental study, to be conducted at lower incident beam energy as well as improving the modelling of phase contrast imaging by using Monte Carlo techniques.

Digital breast tomosynthesis (DBT) reduces breast tissue overlap, which is a major limitation of digital mammography. However, DBT does not show significant improvement in calcification detection, because of the limited angle and small number of projections used to reconstruct the 3D breast volume. Virtual clinical trials (VCTs) were used to evaluate the benefits of computed super resolution (SR) and the optimal combination of the acquisition parameters to improve calcification detection in DBT. We simulated calcifications that were embedded into software breast phantoms. DBT projections of the breast phantoms with and without calcifications were synthesized. We simulated detector elements of 0.085 mm and reconstructed DBT images using 0.0425 mm and 0.085 mm voxels. Channelized Hotelling observers (CHOs) were trained and tested to simulate five virtual readers. Differences in area under the curve (AUC) between SR images and images synthesized with 0.085 mm voxels were calculated using the one-shot multiple-reader multiple-case receiver operator curve (MRMC ROC) methods. Our results show that the differences in AUC is approximately 0.10, 0.03 and 0.03 for DBT images simulated using calcifications sizes 0.001 mm³, 0.002 mm³, and 0.003 mm³, respectively. SR shows a substantial improvement for calcification detection in DBT. The impact of SR on calcification detection is more prominent for small calcifications.

A number of different physical and digital anthropomorphic breast phantoms have been proposed to assess and optimize the performance of breast x-ray imaging systems. All mimic, to some extent, different characteristics of the breast but a systematic realism of phantom realism applied to a number of phantoms using human readers has not been performed, for either full field digital mammography (FFDM), or digital breast tomosynthesis (DBT). We present a reader study in which radiologists performed a subjective evaluation of the visual realism between a selected group of available software phantoms (Stochastic Solid Breast Texture (SSBT) and power law noise texture), physical phantoms (CIRS BR3D breast imaging phantom and the L1 phantom) and clinical mammography images. Regions of interest (ROIs) of 2×2 cm² and 2×2×3 cm³ , for FFDM and DBT stacks respectively, were scored. The readers were asked to judge how well the ROIs represented real breast texture using a 5-point rating scale. Observer ratings were analysed using the receiver operating characteristic (ROC) methodology and the area under the ROC curve (AUC) was used as the figure-of-merit (FOM). The Mann-Whitney test was used to assess the differences between separate groups. For the question of breast texture realism, the SSBT and power-law noise texture images obtained a high score. For DBT, SSBT was also found to have a high visual realism while the power-law noise texture images were found to have mediocre visual realism.

In silico reproductions of clinical exams represent an alternative strategy in the research and development of medical devices, which permit to avoid issues and costs related to clinical trials on patient population. In this work, we present a platform for virtual clinical trials in 2D and 3D x-ray breast imaging. The platform, developed by the medical physics team at University of Naples, Italy, permits to simulate digital mammography (DM), digital breast tomosynthesis (DBT) and CT dedicated to the breast (BCT) examinations. It relies on Monte Carlo simulations based on Geant4 toolkit and adopts digital models of patients derived from high-resolution 3D clinical breast images acquired at UC Davis, USA. Uncompressed digital breast models for BCT exam simulations were produced by means of a tissue classification algorithm; the compressed digital breast models for simulating DM and DBT are derived by the uncompressed ones via a simulated tissue compression. For a selected exam, specifications and digital patient, the platform computes breast image projections and glandular dose maps within the organ. Energy integrating a well as photon counting and spectral imaging detection scheme have been simulated. The current version of the software uses the Geant4 standard physics list Option4 and simulates and tracks <10⁵ photons/s, when run on a 16-core CPU at 3.0 GHz. The developed platform will be an invaluable tool for R and D of apparatuses, and it will permit the access to clinical-like data to a broad research community. Digital patient exposures with the available phantom dataset will be possible for the same patient-derived phantom in uncompressed or compressed format, in DM, DBT and BCT modalities.

Virtual Clinical Trials (VCT) are an effective tool to evaluate the performance of novel imaging systems using computer simulations. VCT results depend on the selection of virtual patient populations. In the case of breast imaging, virtual patients should be matched to a desired clinical population in terms of selected anatomical or demographic descriptors. We are developing a virtual population of women who participated in the Malmö Breast Tomosynthesis Screening Trial (MBTST). We have used clinical values of the compressed breast thickness and volumetric breast density to develop a multidimensional distribution of women in MBTST. Breast density and thickness values were obtained from anonymized, previously collected tomosynthesis images of 14,746 women. In this paper, we compare several approaches to identify clinical subpopulations and select virtual patients that represent various groups of clinical subjects. We performed two methods to identify clinical subpopulations by clustering clinical data using the K-means algorithm or woman’s age. The obtained clusters have been explored and compared using the silhouette mean. The K-means algorithm yielded grouping of MBTST data into two clusters; however, that grouping was, shown to be suboptimal by the silhouette analysis. The agebased clustering showed significant overlap in terms of breast thickness and density. We also compared two approaches to select sets of representative phantoms. Our analysis has emphasized benefits and limitations of different clustering methods. The preferred method depends on the specific task that should be addressed using VCTs. Simulation of representative phantoms is ongoing. Potential correlations with pathological findings and/or parenchymal properties will be investigated.

Digital breast tomosynthesis (DBT) systems utilize an isocentric acquisition geometry which introduces imaging artifacts that are deleterious to image reconstructions. The next-generation tomosynthesis (NGT) prototype was designed to incorporate various x-ray source and detector motions for the purpose of investigating alternative acquisition geometries for DBT. Non-isocentric acquisition geometries, acquisitions that vary the image magnification between projection images, are capable of ameliorating aliasing and other artifacts that are intrinsic to conventional DBT. We used virtual clinical trials (VCTs) to develop custom acquisition geometries for the NGT prototype. A high-resolution (5μm voxel size) star pattern test object was simulated to compare the high-frequency performance of isocentric with non-isocentric image reconstructions. A tilted bar pattern test object was also simulated to compare multiplanar reconstructions (MPR) of isocentric and non-isocentric acquisition geometries. Two source- and detector-motion paths were simulated to obtain super-sampled image reconstructions of the test objects. An aliasing-sensitive metric was used to evaluate spatial resolution performance for two orthogonal frequency orientations. Pairwise comparisons were made for the two frequency orientations between the isocentric and non-isocentric acquisition geometries. Non-isocentric acquisition geometries show an improvement over isocentric acquisition geometries. The greatest improvement was 75.2% for frequencies aligned perpendicular to x-ray source motion, which is the direction of frequencies for which DBT is prone to aliasing. Both frequency orientations exhibit super resolution for non-isocentric geometries. MPR of the tilted bar pattern show zdependent degener

This paper presents a new high-sensitivity PET geometry for high fidelity MRI-compatible PET breast imaging which can scan both breasts simultaneously and have: high sensitivity and resolution; compatibility with MR-breast imaged volume; complete visualization of both breasts, mediastinum and axilla; and a modular design. Whereas contemporary dedicated x-ray and molecular breast imaging devices only scan one breast at a time, this approach relies on an unconventional PET geometry, and is able to provide a PET field of view (FOV) larger than that from dedicated breast MRI. The system geometry is evaluated with GATE Monte Carlo simulations of intrinsic system parameters. Various sized lesions (4-6mm) having [6:1 to 4:1] lesion:background radioactivity ratios mimicking different biological uptake are simulated, strategically located throughout a volumetric anthropomorphic torso. Dedicated breast PET (dbPET) imaging is compared with contemporary clinical PET. The dbPET system sensitivity is >6X greater than for contemporary whole-body PET. The novel, non-conventional system geometry allows for simultaneous dual-breast imaging, along with full medial and axillary imaging. Iteratively reconstructed full-volumetric images illustrate sharper visualization of 4mm lower uptake [4:1] lesions throughout the FOV compared with clinical PET. Image overlap between dedicated breast PET and MRI FOVs is excellent. Simulation results indicate clear superiority over conventional, high-sensitivity whole-body PET systems, as well as improved sensitivity over single-breast dbPET systems. This proposed system potentially facilitates both early detection and diagnosis, especially by increasing specificity of MRI, as well as visualizing tissue heterogeneity, monitoring therapeutic efficacy, and detecting breast cancer recurrence throughout the entire mediastinum.

Early diagnosis of breast cancer can increase survival rate. The assessment process for breast screening follows a triple assessment model: appropriate imaging, clinical assessment and biopsy. Retrieving prior cases with similar cancer symptoms could be used to circumvent incompatibilities in breast cancer grading. Abnormal mass lesions in breast are often co-located with normal tissue, which makes it difficult to describe the whole image with a single binary code. Therefore, we propose an AI-based method to describe mass lesions in semantic abstracts/codes. These codes are used in a searching based method to retrieve similar cases in the archive. This simple and effective network is used for unifying classification and retrieval in a single learning process, while enforcing similar lesion types to have similar semantic codes in a compact form. An advantage of this approach is its scalability to large-scale image retrievals.

Semantic segmentation of breast images is typically performed as a preprocessing step for breast cancer detection by Computer Aided Diagnosis (CAD) systems. While most literature on region segmentation is based on conventional techniques like line estimation, thresholding and atlas-based approaches, such methods may have problems with generalisation. This paper investigates a robust multi-vendor breast region segmentation system for full field digital mammograms (FFDM) and digital breast tomography (DBT) using a U-Net neural network. Additionally, the effect of adding attention gates to the U-Net architecture was analysed. The proposed networks were trained and tested in a cross-validation setting on in-house FFDM/DBT data and the public INbreast datasets, comprising over 10,000 FFDM and 3,500 DBT images from five different vendors. Dice scores were obtained in the range 0.978 - 0.985, with slightly higher scores for the architecture that includes attention gates.

For women of high risk ($>25%$ lifetime risk) for developing Breast Cancer combination screening of mammography and magnetic resonance imaging (MRI) is recommended. Risk stratification is based on current modeling tools for risk assessment. However, adding additional radiological features may improve AUC. To validate tissue features in MRI requires large scale epidemiological studies across health centres. Therefore it is essential to have a robust, fully automated segmentation method. This presents a challenge of imaging domain adaptation in deep learning. Here, we present a breast segmentation pipeline that uses multiple UNet segmentation models trained on different image types. We use Monte-Carlo Dropout to measure each model's uncertainty allowing the most appropriate model to be selected when the image domain is unknown. We show our pipeline achieves a dice similarity average of 0.78 for fibroglandular tissue segmentation and has good adherence to radiologist assessment.

Purpose: To calculate continuous breast density measures from processed images using deep learning. Method: Processed and unprocessed mammograms were collected for 3251 women attending the UK NHS Breast Screening Programme (NHSBSP). The breast density measures investigated included volumetric breast density, fibroglandular volume and breast volume. The ground truth for these measures was calculated using Volpara software on unprocessed mammograms. A deep learning model was trained and validated to predict each breast density measure. The performance of the deep learning model was assessed using a hold-out test set. Results: The breast volume and fibroglandular volume predicted with deep learning were strongly correlated with the ground truth (r=0.96 and r=0.88 respectively). The volumetric breast density had a Pearson correlation coefficient of 0.90. Conclusions: It is possible to predict volumetric breast density from processed images using deep learning.

Estimation of breast density for cancer risk prediction is generally achieved by analysis of full-field digital mammograms. Conventional digital mammography should be avoided if possible in young women because of concerns about potential cancer induction, particularly in those with dense breasts who receive higher doses. This precludes repeated examinations over a short timescale to assess density change. We assess whether density can be accurately estimated in low dose mammograms with one-tenth of the standard dose, with the aim of providing a safe and effective method for use in younger women which is suitable for serial density measurement. We present analysis of data from an on-going clinical trial in which both standard and low dose mammograms are acquired under the same compression. We used both an existing convolutional neural network model designed to estimate breast density and a new model developed using a transfer learning approach. We then applied three methods to estimate density on the low dose mammograms: training on a different mammogram dataset; using simulated low dose data; and training directly on low dose mammograms using cross-validation. Pearson correlation coefficients between measurements on full dose and low dose mammograms ranged from 0.92 to 0.98 with the root mean squared error ranging between 3.37 and 7.27. Our results indicate that accurate density measurements can be made using low dose mammograms.

This work addresses equalization and thickness estimation of breast periphery in digital breast tomosynthesis (DBT). Breast compression in DBT would lead to a relatively uniform thickness at inner breast but not at the periphery. Proper peripheral enhancement or thickness correction is needed for diagnostic convenience and for accurate volumetric breast density estimation. Such correction methods have been developed albeit with several shortcomings. We present a thickness correction method based on a supervised learning scheme with a convolutional neural network (CNN), which is one of the widely-used deep learning structures, to improve the pixel value of the peripheral region. The network was successfully trained and showed a robust and satisfactory performance in our numerical phantom study.

Purpose: The purpose of this study is to test the applicability of a previously developed model observer for detection of calcification clusters in digital breast tomosynthesis (DBT) on five different types of DBT scanners. Methods: A physical phantom with anthropomorphic features (“L1”) was scanned on DBT scanners from five different vendors: Fujifilm Amulet Innovality, GE Senographe Pristina, IMS Giotto Class, Hologic 3Dimensions and Siemens Mammomat Revelation at three dose levels. The phantom images were then prepared for four-alternative forced choice (4AFC) reading study, where six medical physicist observers participated. The image datasets were read as well by a previously developed channelized Hotelling model observer (CHO), also using a 4AFC paradigm, and compared with the human observer results. Results: The percentage of correctly detected calcification clusters (PC) for each target size, dose level and DBT system was compared between the model and the human observers. The goodness of the fit criteria had correlation coefficients varying from 0.94 to 1.00; linear regression slopes ranged from 0.96 to 1.37 and the mean error was between -2.2PC to 5.2PC. Conclusion: The two-step CHO algorithm results closely matched the detectability results of the human observers and can therefore be used for future image quality scoring of the L1 phantom images on these DBT system within the studied dose rate.

Purpose: The methods used in the NHS Breast Screening Programme (NHSBSP) evaluation reports have been consistent over the last 15 years and so it is of value to examine the long term trends in image quality and radiation dose. Method: The mean glandular dose (MGD) measured using 45 mm thick polymethyl methacrylate and threshold gold thickness (t_T) measured using a CDMAM phantom were collated from 30 technical NHSBSP evaluations reports of digital radiography (DR) mammography systems and computed radiography (CR) detectors. Results: There was a wide range of MGD measured for the different systems, but on average there was no significant change in MGD over time for DR. The measured t_T for the 0.1 mm detail sizes significantly improved over time for DR. The MGD required for a system to reach the achievable level of t_T (MGDT) was calculated. This showed that modern DR systems required a lower MGDT. Theoretically all of the DR systems and one CR detector could have been set up to be at the achievable level or better. The MGD, t_T and MGDT results for CR detectors were on average worse than for DR. Conclusions: The quality of the DR systems has improved over time, it was also noted that the difference between the worst and the best system has been decreasing. It would appear that generally manufacturers have set up their systems to take advantage of the improved quality of their systems by improving the quality of the images rather than reducing MGD.

The aim of the study is to compare the small-size objects detection of synthetic 2D (2D-S), breast tomosynthesis (DBT) and standard full-field digital mammography (FFDM) images for different mammography systems. 8 mammography systems of 6 different models and vendors were compared using a home-made phantom composed of CDMAM (Artinis), homogeneous PMMA slabs and BR3D tissue-equivalent slabs (Cirs) containing an anatomic noise. In this study we performed a 4AFC study relative to details with diameters ranging from 100 to 500 μm. We simulate different phantom configurations equivalent to small, standard and large breast adding homogeneous PMMA slabs in order to obtain 3 different equivalent thicknesses. Each phantom configuration was imaged with 2 different dose levels. The processed images were successively divided into squared sub-images, each containing a detail of different size and contrast. The sub-images were rotated and displayed in random order on a 5MP calibrated monitor to 4 trained readers. A percentage correct PC relative each detail was calculated for each mammography, modality and configuration. Preliminary results obtained for configuration A (small breast) and “fixed” dose show that FFDM and DBT images have a significant better detection rates respect to 2D-S for details in the whole range 100-500 μm (p<0.01) while no significant difference in detection for FFDM and DBT was found (p>0.06). Besides these general results, some different behaviours among mammography systems were found.

Purpose: To investigate the effect of age of mammography systems on the number of reported faults and equipment downtime. Method: Each screening unit in the NHS breast screening programmes logged any equipment faults in a centralised online fault reporting database. Data on faults occurring in 2018 for digital mammography systems were analysed. The relationship between the age of mammography systems and the number and consequences of equipment faults was examined. Severity of consequences was reported as number of days downtime, and number of cancelled appointments. Results: 2141 reported faults on 432 mammography systems were included in this analysis. The average age of an x-ray set at the time when a fault occurred was 5.76 years. 72% of mammography systems experienced five or fewer faults in 2018. A significant increase was observed in the number of faults and days of downtime reported on mammography systems six years old or greater, compared to those which were five years old or younger at the time the fault occurred (p<0.05 one-tailed, two-sample t-Test assuming unequal variances). A few mammography systems experienced a high number of faults and days downtime within their first year. The highest average number of reported faults and the highest severity of the consequences were found for mammography systems of around nine years old. Conclusions: The data presented could be used as evidence to support guidance on the age at which mammography systems used for breast screening should be replaced.

The physical performance metrics of computed-radiology (CR) systems used in screening mammography are lower than those of digital-radiology (DR). Also, the lack of quality assurance procedures in some countries might have a technology-dependent impact on image quality and dose. The Mexican Secretary of Health owns over 300 mammography units for breast cancer screening, about half of them of CR technology. We´ve investigated the performance of 20 CR and 4 DR units in 13 Mexican States, applying over 30 quality-control tests associated with general equipment performance, X-ray source, automatic exposure control, mean glandular dose (MGD), image receptor and image quality, and display conditions. Tests were applied following international protocols and their compliance criteria. None of the systems passed all the significative tests. For CRs, the worst performance was observed in compensation for breast thickness, signal-to-noise ratio (SNR) homogeneity, CDMAM thresholdthickness, sensitivity matching of CR plates, and the presence of artefacts. The worse performance of DRs was in compression force, SNR homogeneity, and artefacts. MGD values agreed with recommendations for 2-7 cm PMMA thickness in 50% of CRs and 75% of DRs. The dominance of quantum noise over other components was evaluated by 4 criteria endorsed by different organizations, and results depended on the applied criterium. Analysis of the maintenance procedures suggested that one explanation for these poor results might be the complex CR technology, where the x-ray generation is controlled by a unit fabricated by one manufacturer and the image generation occurs in a non-integrated unit from a different manufacturer.

The aim of this work was to validate an innovative and simplified method for threshold contrast evaluation in digital mammography based on a single exposure of a homemade test tool. A homogeneous region was used to calculate threshold contrast for details of varying size with a statistical approach. An included aluminium step wedge permitted to express the threshold contrast also as an absolute quantity in terms of mm Al. The method has been applied to 9 mammography systems from 5 vendors with different acquisition geometries and detector characteristics in a variety of exposure setups. MTF data were used to correct statistical threshold contrast results for different spatial resolution properties. Validation of the statistical method was carried out as comparison with the widespread method of CDMAM image analysis. Threshold contrasts LC_th obtained with the two methods showed good correspondence. For all systems the ratio, averaged over all exposure conditions, laid within ± 15% from the mean. With respect to the overall figure of merit IQF_inv a linear correlation between the two methods was demonstrated in every unit (r > 0.94 units 1-8, r = 0.89 unit 9). The ratio between the IQF_inv values determined with CDMAM analysis and the statistical analysis was calculated for every exposure condition and averaged for each system. Theses average ratios laid within ± 10% from the mean IQF_inv determined from all units. This is equivalent to a universal angular coefficient between the IQF_inv determined with the two methods, valid for all systems. Finally, mean LC_th ratio values calculated from all data were used to propose reference values for acceptable and achievable threshold contrasts LC_th expressed in terms of mm of aluminium, in analogy to threshold contrasts in terms of gold thickness for CDMAM evaluation.

Anthropomorphic breast phantoms allow realistic testing of mammography systems covering the entire imaging chain and can be used for detectability studies. However, the phantoms usually simulate a specific anatomical situation. If the detectable lesions are modelled directly into the phantoms, the creation of an image database for observational studies in humans requires numerous phantoms. The possibility of flexibly inserting lesions at variable positions within the phantom would considerably reduce the number of phantoms required and thus the manufacturing costs. The aim of this study was to develop a concept for adding simulated microcalcification clusters to our anthropomorphic phantoms, using a 3D printed base plate with movable inserts and crushed eggshells to simulate microcalcification. Mammograms were taken with the base plate under the corresponding phantom for different cluster positions. The resulting mammograms show that the microcalcification clusters overlap the anatomical structures simulated by the anthropomorphic phantom at different positions, while the 3D printed base plate and inserts are not visible. The developed concept may facilitate the provision of a set of images for system tests, including impact of postprocessing on diagnostic image quality, image databases for observational studies and education.

The aim of this work is to estimate the percentage depth dose (PDD) and the mean glandular dose (MGD) at anthropomorphic breast phantoms using calibrated TLDs. For this task the TLD-100H was initially selected and calibrated in terms of air kerma, using a Radcal ionization chamber (IC). The experimental procedure was performed at a mammograph Mammomat 3000 Nova, located at the CDTN/CNEN facilities and all the detectors were exposed with 28 kV with distinct anode/filter combination (Mo/Mo, Mo/Rh and W/Rh). Furthermore, the TLDs were placed at the surface of anthropomorphic compressed breast phantoms (36-85 mm) and at 1,2 cm for depth doses measures, for PDD analyses. The MGD were estimated from entrance surface doses and using Dance´s method. The dose–response curves for the TLDs indicated a good correlation coefficient (R² = 0,99) for all anode-filter combination with an uncertainty lower than 17%. The uncertainties of the measurements increased to maximum 35 % when the TLDs are placed the anthropomorphic breast phantom. The PDD were maximum 70%, at 1,2 cm, for the Mo/Mo target-filter combination. Regardless the target/filter (T/F) combination, the TL responses and consequently the MGDs substantially increased with the breast thickness. A maximum MGD of approximately 4.5 mGy was estimated for the 85 mm thickness breast, exposed to the Mo/Mo combination.

Evaluation of the performance of all mammographic equipment requires the imaging and interpretation of test objects or phantoms and risk evaluation. The Mean glandular dose (MGD) is the relevant quantity for dosimetric regulatory actions in screening mammography and its evaluated usually obtained with a polymethylmethacrylate (PMMA) phantom. This study evaluated by spectrometry technique the linear attenuation coefficients by transmitted intensity of commercially tissue-equivalent phantom for mammography, Computerized Imaging Reference System (CIRS) with different glandularity-adipose content (0-100, 30-70, 50-50, 70-70 and 100-0) % and PMMA. The measurements were realized with x-ray beams produced by Mo/Mo, Mo/Rh and W/Rh. The results are available for the x-ray characteristic photo-peaks (k_α1 Mo 17.48 keV and k_α2 Mo 19.61 keV). For spectrometry measurements use an AMPTEK XR-100T Cadmium Telluride (CdTe) detector. Simulations using the MCNPx Monte Carlo code were performed to evaluate the detector response. The results of corrections performed using the methodology presented here were compared with spectra corrected by other authors, for validation purposes. The comparisons showed that this methodology was adequate for the correction of spectra from 2 keV to 200 keV. The results of linear attenuation coefficients here were compared with experimental and theoretical values that were previously reported. In this work we show the adequacy of the results found for PMMA and CIRS 50-50 compared. In addition, it brings the evaluation of the attenuation coefficient for kα2 Mo with 19.61 keV for all CIRS combinations.

The use of phantoms is the most appropriate way to guarantee a control and quality assurance program for images formed from the interaction of ionizing radiation with matter. However, the existence of anthropomorphic phantoms for the evaluation of images obtained through the breast tomosynthesis technique are scarce, do not meet all the needs of a quality control program and are difficult to obtain, mainly due to the way they should be made and the availability of materials capable of simulating structures typical of human tissues. This paper analyzes the possibility of using commercially available photopolymers to be used in high-quality 3D printers for the production of an anthropomorphic breast phantom, with structures similar to the skin, blood, adipose tissue, and gland tissue that meet the peculiarities of tomosynthesis exams. We investigated five photopolymers already printed in the form of solid parallelepipeds, each with four different thicknesses, concerning their mass attenuation coefficients and their half-value layer. Two of the five samples interacted with x-rays in the same way as x-rays interact with adipose and glandular tissues, suggesting that they are apt to be readily used to simulate such tissues. We suggest investigating other photopolymers, printing a full-scale phantom, and partnering with researchers and developers of 3D printing supplies as a way to achieve a new and improved class of anthropomorphic phantoms.

Aim: In recent years Phase Contrast Tomography (PCT) has been rapidly progressing towards clinical translation as an advanced imaging technology for breast cancer diagnosis. Recent optimization of PCT with mastectomy samples has refined imaging protocols and biomedical-engineering prowess is now required to formalize patient table and breast immobilisation requirements. PCT imaging requires women to lie in prone position similar to conventional breast CT, however the imaging couch rotates above the beam allowing exposure of the breast beneath. Motion artefact through involuntary movement of the breast through the rotation cycle has the potential to reduce diagnostic quality of the results. Methods: This paper details the biomedical engineering cycle of breast holder development alongside medical physics considerations. Breast immobilisation via a plastic or silicone supporting material which is sufficiently transparent for X-rays in the targeted energy range is explained, including the two step process of considering single cup versus double cup solutions and how mild-suction to the breast can be implemented in order to maximum breast tissue visualization and assist with dose uniformity. Results: Considering patient comfort, breast positioning and implications upon attenuation and phase shift, a number of models were developed in Australia and Italy. Early prototypes are described here with some preliminary imaging. Considerable work is taking place over the next three months as models undergo imaging with mastectomy samples at the Imaging and Medical Beamline at the Australian Synchrotron and the ELETTRA Synchrotron Italy. Consumer representatives will be rating the immobilisation device for comfort prior to the start of clinical trials in 2020.

X-ray phase sensitive imaging has been employed in the preclinical settings for more than two decades. The advancement in the technology has allowed to potentially translate this innovative imaging technique to the clinical environment. In-line phase sensitive imaging technique has shown promising potential to be used for breast cancer imaging. A high energy phase sensitive breast tomosynthesis (PBT) prototype system based on the inline phase sensitive imaging technique has been developed for the potential imaging in clinical environment. The prototype system incorporates a microfocus x-ray tube and a flat panel detector having a pixel pitch of 70μm. The microfocus x-ray tube has a tungsten (W) anode, Beryllium (Be) output window and a focal spot size that ranges from 18-50μm, depending on the output power. The x-ray tube/detector configuration produces a geometric magnification (M) of 2.2 and acquires 9 projection views within 15 degrees or 30 projection views within 30 degrees in stop-andshoot scanning mode. A single distance phase retrieval scheme method based on the Phase-Attenuation Duality (PAD) principle is applied on the angular projection views. A filtered back-projection operation reconstructs a set of tomogram slices at 1mm incremental depth within the breast along the z-direction. American College of Radiology phantom images demonstrate that both 2D and tomosynthesis images acquired on the prototype system meet the minimum criteria set by the Mammography Quality Standard Act. We have also imaged mastectomy specimens with the PBT prototype system at the University of Utah Huntsman Cancer Hospital. PBT 2D images and tomosynthesis images slices demonstrate image quality comparable to a conventional digital breast tomosynthesis clinical system.

In this work we present the BiPlanar Breast PET (BBPET) system. It is composed of two movable paddles than can be placed in several configurations in order to allow the imaging of the breast and the pectoral wall. Each paddle is formed by several detectors formed by continuous crystals, coupled to SiPM arrays. The system has been evaluated measuring the spatial resolution, NECR, uniformity and contrast coefficients. The results show a transaxial resolution of 1.5mm, a NECR peak of 319kcps at 12.6MBq, a uniformity of 10% and recovery coefficients >0.7 for the 3, 4 and 5mm rods. The system has good imaging capabilities for imaging patients in the clinical routine.

In digital mammogram visualization, text is usually among the brightest objects. When bright text is near an area of interest in the tissue image, it can annoy the readers and may reduce the perception of tissue details. To mitigate this effect, we propose segmentation of the frame buffer to annotation text and mammographic image using artificial intelligence. An appropriate luminance can be re-assigned to each area (i.e., dim annotations and/or brighten mammographic image to the maximum luminance). Existing text detection and/or segmentation tools we tested did not work for this purpose because they produce false positives (i.e., parts of breast are detected as text and dimmed). That is perhaps because such methods were designed (or trained to, in case of deep learning methods) for natural images. We investigated two state-of-the-art segmentation architectures DeepLabV3+ and Mask R-CNN, as well as a “shallow” text detection method based on maximally stable external regions (MSER). We generate the training data by adding random text to the background of publicly available mammographic images. DeepLabV3+ trained to our data produced promising results while Mask R-CNN and MSER did not. Keywords: Convolutional neural

Receiver operating characteristic (ROC) examination was performed to investigate the effectiveness of high-luminance monitors in digital X-ray mammography. For this purpose, an original breast phantom consisting of adipose and fibroglandular equivalent tissues with an identical X-ray absorption characteristic over the entire mammographic photon energy range was developed. Furthermore, the phantom’s fibroglandular density and distribution could be changed arbitrarily. Three types of lesions, microcalcification, mass, and spiculated, were inserted into the breast phantom, and the ROC examination was performed by five radiological technologists certified in screening mammography, to obtain the area under the curve. A liquid crystal display (LCD) monitor with 5 megapixels in a 21-inch display size calibrated to a grayscale standard display function curve was used for the observation. The monitor was set to 600, 900, and 1200 cd/m² in maximum luminance. The experimental details were fibroglandular density of 25%, respective 50 positive and negative images, and free observation time and distance. As a result, the dependence on monitor luminance differed according to the lesion type. The detectability of microcalcification increased with the increase in the luminance of the monitor. Spiculated lesions were similar for all luminance changes. The detectability of mass lesions was significantly higher at 900 cd/m² than at 600 cd/m² . There was no significant difference between those at 900 cd/m² and 1200 cd/m² . In conclusion, the maximum luminance of the diagnostic LCD monitor for mammography should be at least 900 cd/m² to guarantee stable detectability.

In recent years, convolutional neural networks (CNN) have found increasingly active application in the field of computer-aided diagnosis (CAD) research. Typically, general-use, high-performance detectors are designed using machine learning, the training of which is conducted by applying comprehensive sets of case images having various variations. In this study, we show that, when configuring CNN training data, dividing the data into multiple subsets and adjusting their ratios, instead of providing the data uniformly, has the potential for effective learning. We propose in this study a learning method by which CNN learning using these subsets is incrementally repeated. In this study, subsets of breast cancer mass learning data based on mass size and intensity were created. Using multiple data sets prepared for use in the evaluation of a CNN that had been subjected to learning, optimal ratios were considered and, based on this, performance evaluations using actual unknown data were conducted. Next, the ratios of evaluation data subsets having numerous detection errors were raised and relearning conducted. This process was repeated multiple times, as long as increases in the area under curve (AUC) were observed, thus enabling the design of a high-performance CNN. As a result of applying unknown data to this CNN, we found that it exhibited a higher AUC than a CNN to which learning data was simply provided comprehensively, demonstrating the effectiveness of the proposed learning method.

This study is aimed to automatically segment mammary gland region into scattered mammary glands and fatty breasts using deep learning method. Total 433 mediolateral oblique-view mammograms of Japanese women were collected and confirmed for scattered mammary glands or fatty breasts; using BI-RADS’s classification. First, manually contoured mammary gland regions were determined for all mammograms as ground truths by three certified radiological technologists. Second, the U-net model was employed to segment the mammary gland region automatically. This model is a type of convolutional neural network (CNN) mainly aimed at medical image segmentation. The segmentation accuracies were assessed based on five criteria, Dice coefficients, breast densities, mean gray values, centroids, and sizes of mammary gland region. The Dice coefficient was 0.915. The mean size of mammary gland regions obtained by the Unet was 8.7% larger than that of the ground truths. The mean centroid coordinates of mammary gland regions by the U-net were shifted 1.6 and 5.4 mm on average in mediolateral and craniocaudal directions, respectively from ground truths. The mean gray value of mammary gland regions obtained by the U-net was only 0.4% higher compared with ground truths. The resultant difference was 0.4% on average in breast densities between ground truths and the segmented mammary gland regions. We found significant similarity in the ground truths and the data generated by deep learning on all the parameters, thereby attesting the efficacy of this method for segmenting the mammary gland regions of not only the dense breasts but also the scattered mammary gland- and fatty- breasts.

Knowledge of the compressed breast shape can be valuable information to improve tomosynthesis reconstructions. The goal of this work was to use a convolutional neural network to refine the shape as estimated from tomosynthesis projection data. Training data was created by generating random three-dimensional breast shapes and simulating the limited angle projections. A rough approximation of the breast shape was made by segmenting and then back-projecting the projection data. Following this, a 3-layer u-net was trained on 900 pairs of simulated breast shapes and the corresponding shape estimates. The resulting network was applied to 100 test cases, where it significantly reduced the average distance between the surfaces of the true and estimated breast shapes from 2.3 mm to 0.5 mm (p < 0.001). If these results can be confirmed using patient data, it is likely that advanced image processing techniques that rely on precise knowledge of the compressed breast shape will become feasible since our work now provides a method to obtain such an accurate estimate of the breast shape without the need for any additional imaging hardware.

Cribriform architecture is a histological pattern reminiscent of Swiss cheese that is commonly recognized in ductal carcinoma in situ (DCIS) of the breast observed by microscope. However, there are only a few three-dimensional studies to elucidate whether each glandular cavities of cribriform pattern are connected or not. The main reason for paucity of three-dimensional studies is that the conventional reconstruction based on histological sections requires laborious and time-consuming works. In this research, we first performed three-dimensional reconstruction of the cribriform pattern using crystal analyzer-based phase contrast technique, X-ray dark field computed tomography (XDFI-CT), which provides high contrast image of biological soft tissue with non-destructive and non-staining approach. Then, we propose a machine-learning-based method to extract the cavity from XDFI-CT images. Finally, we show that the useful information to analyze the cribriform patterns in DCIS such as the density and volume of the cavity can be obtained from the XDFI-CT images.

It can be difficult for clinicians to correctly determine biopsy or follow-up for masses on breast ultrasonographic images. The purpose of this study was to develop a computerized determination scheme for histological classification of masses using a combination of CNN (convolutional neural network) features and morphologic features. The database consisted of 585 breast ultrasonographic images. It included 288 malignant masses (218 invasive carcinomas and 70 noninvasive carcinomas) and 297 benign masses (182 fibroadenomas and 115 cysts). In the proposed method, CNN features and morphologic features were first determined from a mass. The CNN features were defined by reducing the dimensionality of the output of the final pooling layer in GoogLeNet using a principal component analysis. The morphologic features were also defined by taking into account image features commonly used for describing masses on breast ultrasonographic images. A support vector machine (SVM) with the CNN features and the morphologic features was employed to classify among histological classifications of masses. Three-fold cross validation method was used for training and testing the GoogLeNet and the SVM. The classification accuracies with the proposed method were 84.4% (184/218) for invasive carcinomas, 72.9% (51/70) for noninvasive carcinomas, 85.7% (156/182) for fibroadenomas, and 87.8% (101/115) for cysts, respectively. The sensitivity and the specificity were 87.2% (251/288) and 93.3% (277/297), whereas the positive predictive value and the negative predictive value were 92.6% (251/271) and 88.2% (277/314). The proposed method yielding high classification accuracies would be useful in the differential diagnosis of masses on ultrasonographic images as diagnosis aid.

Breast magnetic resonance imaging (MRI) has a higher sensitivity of early breast cancer than mammography, but the specificity is lower. In MRI examination at clinical practice, multiple MRI sequences are usually acquired to achieve high diagnostic accuracy. The purpose of this study was to develop a computerized classification scheme for distinguishing between benign and malignant masses by integrally analyzing multiple MRI sequences with convolutional neural networks (CNNs). Our database consisted of four MRI sequences for 43 patients with masses. It included T1-weighted images, T2- weighted images, dynamic contrast material-enhanced magnetic resonance imaging (DCE-MRI) images, and the difference images of the DCE-MRI images for each patient. In training the CNNs, the CNNs were first trained independently for each MRI sequence. The CNN features extracted from four MRI sequences with the trained CNNs were then inputted to a support vector machine (SVM) for distinguishing between benign and malignant masses. A k-fold cross validation method (k=3) was used for training and testing the CNNs and the SVM. With the proposed method, the classification accuracy, the sensitivity, the specificity, the positive predictive value, and the negative predictive value were 88.4% (38/43), 90.0% (27/30), 84.6% (11/13), 78.6% (11/14), and 93.1% (27/29), respectively. The classification performance with the proposed method analyzing multiple MRI sequences was substantially greater than those with CNNs analyzing one MRI sequence. The proposed method achieved high classification performance and would be useful in differential diagnoses of masses as diagnostic aid.

Artificial intelligence (AI) applications are increasingly seeing use in breast imaging, particularly to assist in or automate the reading of mammograms. Another novel technique is mechanical imaging (MI) which estimates the relative stiffness of suspicious breast abnormalities by measuring the distribution of pressure on the compressed breast. This study investigates the feasibility of combining AI and MI information in breast imaging to provide further diagnostic information. Forty-six women recalled from screening were included in the analysis. Mammograms with findings scored on a suspiciousness scale by an AI tool, and corresponding pressure distributions were collected for each woman. The cases were divided into three groups by diagnosis; biopsy-proven cancer, biopsy-proven benign and non-biopsied, very likely benign. For all three groups, the relative increase of pressure at the location of the finding marked most suspicious by the AI software was recorded. A significant correlation between the relative pressure increase at the AI finding and the AI score was established in the group with cancer (p=0.043), but neither group of healthy women showed such a correlation. This study suggests that AI and MI indicate independent markers for breast cancer. The combination of these two methods has the potential to increase the accuracy of mammography screening, but further research is needed.

We aimed to test two different software based on the deep learning technology versus two senior and one junior radiologist on a recall-based model for mammography. We performed a retrospective, monocentric, multi-reader study in the Centre Hospitalier de Valenciennes in the north of France. A set of examinations from a daily practice, with both screening and diagnostic studies, has been interpreted by 3 radiologists and the two AI based algorithms. The dataset has been enriched with BIRADS 4 and 5 cases in order to have a number of cancer cases sufficient to have statistically significant results. In total, 140 examinations have been included in the final dataset. Sensitivity (true positive rate - TPR), False positive rate (FPR), and recall rate per BI-RADS category were considered as endpoints for each of the radiologists. To compute these metrics all the included cases were considered as positive if the initial BI-RADS was equal or higher than 3 and as negative if the initial BI-RADS was 1 or 2. Additional analysis have been carried out taking into account the biopsy report (if any) as ground truth. While both the algorithms and radiologist have a good and comparative rate of sensitivity and FPR, the test based on BI-RADS categories (i.e. the number of cancer per BI-RADS category), showed heterogeneous results, with bad performances for one of the tested software on the extremes score of BI-RADS. We concluded that one of the analysed software cannot be used in the current clinical practice without further improvements, the second one shows promising results, but other studies are needed to have a robust external validation before being used in a daily practice.

Digital mammography (DM) has been considered as the primary modality for breast cancer screening. The relative amount of breast fibroglandular tissue, referred to as percent breast density (PD), has been considered as an important factor associated with breast cancer. We have developed and tested a robust method to accurately segment the pectoral muscle and the breast area using a deep learning approach. We use a U-Net architecture with a ResNet decoder to increase the depth of features. The architecture is trained using 555 DM images and tested and validated on an independent set of 555 images. The results show that our network achieves an average and standard deviation dice coefficient of 94.86% ± 1.93%, respectively, and sensitivity of 96.31% ± 1.87%. The method present here can be considered as the first step toward the automatic estimation of PD.

The paper proposes a novel approach for the identification of cancerous regions located in a dense part of a breast. This task is particularly challenging even for experienced radiologists due to lack of clear boundaries between the cancerous and normal tissue. Multi-scale analysis of structured micro-patterns generated from local binary patterns (LBP) was used to generate a very small number of features which allowed for successful detection of cancerous regions. The proposed technique was tested on two publicly available datasets: Digital Database for Screening Mammography (DDSM) and INbreast. The area under the receiver operating characteristic (AUC) curve for DDSM with 2 features only was 0.99 and 0.92 for INbreast with 3 features.

Mammographic sensitivity in breasts with higher density has been questioned. Higher breast density is also linked to an increased risk for breast cancer. Even though digital breast tomosynthesis (DBT) offers an attractive solution, for varied reasons it has not yet been widely adopted in screening. An alternative could be to boost the performance of standard mammography by using computer-aided detection based on deep learning, but it remains to be proven how such methods are affected by density. A deep-learning based computer-aided detection program was used to score the suspicion of cancer on a scale of 1 to 10. A set of 13838 mammography screening exams were used. All cases had BIRADS density values available. The set included 2304 exams (11 cancers) in BIRADS 1, 5310 (51 cancers) in BIRADS 2, 4844 (73 cancers) in BIRADS 3 and 1223 (22 cancers) in BIRADS 4. A Kruskal-Wallis analysis of variance showed no statistically significant differences between the cancer risk scores of the density categories for cases diagnosed with cancer (P=0.9225). An identical analysis for cases without cancer, showed significant differences between the density categories (P<0.0001). The results suggest that the risk categorization of the deep-learning software is not affected by density, as though some density categories receive higher risk assessments in general, this does not hold for cancer cases, which show uniformly high risk values despite density. This shows the potential for deep-learning to improve screening sensitivity even for women with high density breasts.

Segmentation of abnormalities in mammographic images has been a challenge due to poor contrast between the abnormality and surrounding area. In general Computer Aided Diagnostic (CAD) systems are composed of three major step: i.e segmentation followed by feature extraction and classification. The ultimate results of CAD systems are strongly influenced by the results of segmentation, as poor segmentation leads to incorrect features being extracted and ultimately incorrect classification. In this paper we propose a novel method for the segmentation of masses in mammographic images by using a simple method based on the information provided in histograms. After applying some preprocessing on the mammographic images in order to smooth the image histograms, optimal threshold values were obtained for each mammogram which can deal with uni-modal and bi-modal histograms. Segmentation results show an improved area for the mammogram masses, representing the true shapes of the abnormalities. We tested our algorithm on 233 benign and 233 malignant abnormalities. Results are in line with state of the art methods.

Retrieval of similar cases can help radiologists in efficient diagnosis, treatment planning, and preparation of reports for new cases. In this study, similarities of pairs of lesions were estimated using convolutional neural networks with subjective similarity data. The network was trained with pairs of mammograms (MG), pairs of ultrasound images (US), and both as input data and the corresponding similarity ratings by expert radiologists as teacher data. Based on the estimated similarity, the cases with the highest similarities were retrieved for a test case. The precisions of selecting pathology-matched relevant cases were compared for the networks using different input data. In this study, rather a simple network architecture, which takes a pair or pairs of input images and has one regression output layer corresponding to the similarity, provided higher precisions. The precisions using mammograms, ultrasound images, and both modalities were 0.72, 0.68, and 0.80, respectively. The highest precision was obtained by the use of one network with multimodality image inputs than combining the outputs by two separate networks for MG and US data. Relatively high precision indicates that the presentation of reference images can be useful for assisting breast cancer diagnosis.

Slice by slice visualization of Digital Breast Tomosynthesis (DBT) data is time consuming and can hamper the interpretation of lesions such as clusters of microcalcifications. With a visualization of the object through multiple angles, 3D volume rendering (VR) provides an intuitive understanding of the underlying data at once. 3D VR may play an important complementary role in breast cancer diagnosis. Transfer functions (TFs) are a critical parameter in VR and finding good TFs is a major challenge. The purpose of this work is to study a methodology to automatically generate TFs that result in appropriate and useful VR visualizations of DBT data. For intensity-based TFs, intensity histograms were used to study possible relationships between statistics and critical intensity values in DBT data. The mean of each histogram has proved to be a valid option to automatically calculate those critical values that define these functions. At this stage, eight visualizations were obtained by combining several opacity/color intensity-based functions. Considering the gradient, ten visualizations were obtained. Nine of the ten TFs were constructed considering the peaks of gradient magnitude histograms. The tenth function was a simple linear ramp. Finally, three intensity-based and three gradient-based functions were selected and simultaneously used. This resulted in nine final VR visualizations taking both information into account. The studied approach allowed an automatic generation of opacity/color TFs based on scalar intensity and gradient magnitude histograms. In this way, the preliminary results obtained with this methodology are very encouraging about creating an adequate visualization of DBT data by VR.

In this study we present a preliminary evaluation of the inter-operator (InterOp) and intra-operator (IntraOp) variability of Quantitative Ultrasound features based on first-order speckle statistics used in breast cancer characterization. Ultrasound echo signals from ten patients with biopsy-confirmed invasive ductal carcinomas were acquired in vivo in the radial and antiradial planes with a commercial ultrasound system using a linear array transducer. Each patient was scanned by three radiologists, each of which performed three acquisitions allowing the patient to reposition in between acquisitions. Parametric images of six QUS features obtained from the first order statistics of the speckle pattern of ultrasound images were computed, and the mean feature value within the lesion boundary was compared between pairs of images from different radiologists or acquisitions from the same radiologist. In general, the InterOp variability was 1.2 times larger than the IntraOp variability. These differences were not significant in the radial plane. In addition, features with similar InterOp and IntraOp variability were the ones with the largest overall variability.

The aim of this study is to evaluate the effect of adding digital breast tomosynthesis (DBT) to digital mammography (DM) on sensitivity and specificity scores for readers with different DM and DBT experience compared with that of DM alone. Ethical committee approval was obtained. 41 DM and DBT cases (22 cancer, 19 normal), each containing two views, were reviewed by 18 readers, divided into groups according to level of experience with DBT and DM. Readers were asked to report each case in two modes (DM and DM+DBT) using a 5-point scale (1- Normal, 2- Benign, 3- Equivocal, 4-Suspicious, 5- Malignant). The radiologists’ diagnostic performance was compared between DM and DM + DBT and evaluated by sensitivity and specificity. Readers with no DBT workshop showed higher sensitivity using DM+DBT compared with DM (P-value 0.03). Female readers, readers with less than 5 years of DM experience, readers with more than 20 mammography reads per week, readers who are not using DBT in clinical practice, readers with mammography fellowship, and readers who had a DBT workshop showed a significantly higher specificity using DM+DBT in comparison to DM alone (P-values 0.01, 0.01, 0.02, 0.03, 0.03, 0.03, 0.01 respectively). The current study showed that the addition of DBT to DM might not significantly change the readers performance in terms sensitivity, however it may result in less number of recalls to additional examinations which provides a substantial benefit in the screening and diagnostic settings.

Objectives: To assess the accuracy of size measurements of spiculated masses for both 2D mammography (MX) and Digital Breast Tomosynthesis (DBT) in comparison to histopathologic outcomes in breast cancer patients. Methods: A retrospective study was conducted from January 2015 till December 2016 with inclusion of biopsy proven breast cancer patients who presented themselves with masses with spicules. Two readers performed size measurements of both the masses only as well as of the masses with spicules. Results were compared to histopathological results. Results: A total of 180 spiculated masses were included in this study. Analysis showed that for the mass only measurements both readers showed a significant underestimation on both 2D Mx as well as DBT versus histopathology. While the mass with spicules measurements showed more variable results, partly depending on tumor size and clinical working experience. We found no influence of viewing direction or breast density. Conclusions: Even with DBT it is very difficult to determine the correct size of spiculated mases. Measurements of only the central mass are not correlating well with histopathology. Especially for larger masses (> 25mm), radiologists tend to underestimate the size of spiculated masses. For smaller masses measurement of the mass center only as well as inclusion of the spicules correlated well with pathology sizes.

Breast compression in mammography is used to obtain suitable imaging. Compression reduces breast thickness and spreads the tissue, leading to improved image quality and reduced radiation dose. Breast compression is, however, uncomfortable for clients, and may often cause strong pain. In this study, we obtained pressure distribution and X-ray images of a compressed breast simulated phantom using a pressure sensor. The purpose of this study was to visualize the pressure distribution and measure how much the phantom was spread by compression. The pressure sensor was placed on a breast support table, and the breast phantom was positioned in the medio-lateral oblique direction. Compression was applied in steps of 20 N from 40- 180 N. Indicated values of the pressure sensor and width of the compressed phantom were measured at each increased force step. Width differences from a non-compressed phantom were calculated at each applied compression force. The pressure distribution showed nonuniform pressure over the breast phantom, and high pressure in the juxta-thoracic region, which decreased toward the periphery. The width of the breast increased with increasing applied compression force in all positions, but the change became smaller after 140 N. By visualizing the pressure, we were able to evaluate the pressure distribution over all of the phantom.

Breast density and lesions characteristic are important determinant to discriminate between benign from malignant tissue. This study was conducted to determine the association between breast density and lesions characteristic with diagnosis error in mammography. A total of 167 mammographic cases irrespective of the indication for mammography, whether screening or diagnostic performed from 2011-2017 were retrospectively identified from the clinical database. Descriptive statistic and chi square test series were used to analyze the data. Majority (89%) mammographic cases in this study had a false positive (FP) and 11% had a false negative (FN) result. FP cases more likely occurred among Malay women (57.6%) as compared to Chinese (33.3%), Indian (6.0%) and others (3.3%). In contrast with FP, FN cases were mostly occurred among Chinese (41.2%) as compared with other races. Masses were the most common type of lesions in both FP (77.3%) and FN cases (88.2%). No significant association was found between breast density and lesions characteristic with diagnosis error in mammography. Most of the lesions found in FP and FN cases were in women with denser breast. Lesions found in FP groups have characteristic that mimic malignancy, while lesions in FN group not necessarily has the criteria of benign lesions.

In this work, images of exams performed in Digital Breast Tomosynthesis (DBT) system were collected retrospectively from 660 Brazilian women, who underwent screening mammography in clinics located in three Brazilian geographic regions. The raw images were processed using the Volpara software, through which the volume and volumetric density of the breast, the contact area between the breast and the tray, the force and pressure of compression and the thickness of the compressed breast were determined. mean breast size was determined by Volpara, resulting in 737 cm³ . The compression force had a median of 79.5 N, with a range of 20 to 160 N and a compression pressure of 9.94 kPa, with a range of 2.6 to 29.5 kPa. The analysis of the correlation between the quantities resulted in percentage dense volume and volume with compression force, r = - 0.259 and r = 0.313, respectively (p < 0.01)), percentage dense volume and volume with compression pressure, r = 0.327 and r = -0.478 (p < 0.01), showing that when it is considered an intrinsic characteristic of the breast there is a greater possibility of standardizing compression through compression pressure instead of compression force.

Mammography and tomosynthesis diagnostic techniques contribute to the patient dose and the estimation of the Mean Glandular Dose (MGD) for x-ray based imaging modalities forms an essential part of quality control (QC) and for system optimization. This study compares three solid‐state dosimeters (SStDs) and one thermoluminescence dosimeter (TLD) with a half value layer (HVL) measurements by ion chamber. These electronics SStDs instruments are very efficient in clinical practice. However, a specialized technician is required for its operation and the clinics do not have this instrumentation. Thus, the application of a TLD system becomes an important means for determining the MGD through the evaluation of K_a,i and HVL. For measurements used a Siemens Mammomat 3000 Nova mammography to evaluate the dosimetry systems. The three electronics SStDs detectors (Accu-Gold Radcal, Unfors RaySafe Xi and Piranha), TLD system and TRS457 table were compared with ion chamber results for HVL. The results showed that SStDs and TLD dosimeters have a flat HVL response under clinical conditions (less than 5%). However, the SStDs detectors tended to overestimate the measurements. However, all results were equivalent (p<0,05) for all HVL evalueted.

The use of mammography equipment attached to a digital breast tomosynthesis (DBT) system is widespread in Japan. Tomosynthesis exposes the breast to X-rays continuously (or in pulses) at different angles in a single scan, yielding multiple projection images. Arbitrary multi-slice images can be reconstructed from the projection images after scanning. However, DBT increases the exposure dose compared to mammography. Therefore, it is necessary to rapidly establish performance evaluation procedures and quality control procedures for DBT. In this study, we conduct quality control measurements on DBT systems sold in Japan by five different companies and examine feasible common items. The purpose of the study was to establish a quality control method for DBT systems in Japan. The measurements were performed based on the EUREF breast tomosynthesis quality control protocol version 1.03. In this study, we attempted to measure 18 items in DBT systems. We examined whether the 18 items could be measured using each system; however, this was not an evaluation of equipment performance based on the measured values. There were some quality control items that were difficult to complete due to the specifications of a DBT system, such as equipment that required pressure during DBT operation, problems due to the shape of the bucky, and equipment that did not have stationary mode. There were also problems related to the availability of measurement data, such as with equipment that could not retrieve projection data and reconstructed data. This study clarified points to be considered for establishing common quality control items. In the future, we will carefully refer to the recently published IEC 61223-3-6, consider international harmonization, and establish DBT guidelines customized for the Japanese market.

The objective of this work is to present the results for quality control tests applied to projection images acquisition in digital mammography and breast tomosynthesis (DBT). Mean glandular doses (MGD) were measured for the examination of series of women and for breast-simulating polymethyl methacrylate phantoms, thus assessing the suitability of the phantoms used for dosimetry in 2D mammography for DBT dosimetry. Moreover, X-ray tube output and half value layer measurements for MGD estimation using phantoms are also presented. Three different mammography/DBT systems were considered in this work: Hologic Selenia Dimensions, General Electric Senoclaire and Pristina and Siemens Inspiration. The results obtained for the different projections were compared with the 2D acquisitions and the differences between the two image modalities were compared.

This study aims to verify the relationship of MGD between four different types of manufacturing mammograms and models and to verify patient characteristic factors and GDM. Using the Volpara software were analyzed a total of 7,000 3D and 2D images. From this analysis were obtained the breast volume density (DVB) and the MGD. Using the DICOM header of the image, we collected the patient's age and compressed breast thickness. The sample of patients presented a mean of 57 (±15) mm of compressed breast thickness(CBT) for the Hologic equipment (range from 19.82 to 100.75 mm) and the medians for the other variables were 51 years (range 25 to 87 years old), 1.75 mGy MGD (0.43 to 4.68 mGy range), and 7.61% DVB (2.16% to 36.89% range). The MGD for GE Senoclaire system and Hologic were higher compared the other evaluated tomosynthesis systems as also higher for MLO projection when compared to CC projection. The Siemens equipment was the system that gave the lowest dose in all breast thicknesses evaluated.

Purpose: Low levels of ghosting are found during quality control of digital mammography (DM). In digital breast tomosynthesis (DBT) mode, the x-ray detector is not cleared during the series of exposures potentially leading to ghosting issues. Methods: The ghosting image factor (GIF) was calculated for systems from three manufacturers as follows. A 45 mm thick polymethyl methacrylate (PMMA) block covering half the detector was imaged, followed by the PMMA block plus 0.1 mm thick aluminium sheet centred on the block, this time positioned to fully cover the detector. This was undertaken for (1) DM and (2) between separate DBT scans. Inter-projection ghosting (3) was measured by moving the PMMA block from covering half the detector to covering the whole detector after the first projection. In method (4), a 3 mm thick aluminium plate with a 0.1 mm thick aluminium sheet on top was placed on the compression paddle at 100 mm above the breast covering half of the detector. Results: The GIF measured via test (1) and (2) was lower than the 0.3 tolerance set in the European guidelines for DM, however higher values were found for one system in DBT mode. Test 3, inter-projection GIF values were between 0.08 and 0.89, mostly above 0.3. Test 4 was a practical method and gave larger GIF values of between 1.07 and 8.01. Conclusions Ghosting factors measured between projections are considerably higher than the GIF measured between separate DBT scans. We propose a simple method using an aluminium plate to estimate the GIF.

Purpose: The purpose of this work is to compare two methods of artifact spread function (ASF) estimation in digital breast tomosynthesis (DBT) and study the feasibility of a task-based ASF estimation. Methods: A homogeneous PMMA phantom with two aluminum spheres with size 0.5mm and 1.0mm was scanned on Siemens Inspiration, Giotto Class and Hologic Dimensions DBT systems. The ASF curves were estimated using a standard method from average pixel values from the DBT planes. A physical phantom with anthropomorphic features, including microcalcification simulating particles of average diameter 0.24 mm and a 3.0mm mass-like lesion, was also scanned on the same DBT systems. The in-focus and out-of-focus planes were read with a newly developed model observer in order to assess target detectability through the different slices and calculate task-based ASF curves for each DBT system. The corresponding curves for the standard sphere and the task-based methods were compared. Results: The ASF curves for the smaller targets, i.e. the 0.24 mm microcalcification particles and 0.5 mm aluminum sphere, were found to match closely, despite the size difference and the ASF curves for the larger targets. The propagation across planes for the 3.0mm mass-like lesion and the 1.0mm aluminum sphere did not match. Conclusions: The task-based ASF estimation gives better clinical relevance to the artifact spread estimation. The ASF of 0.5mm aluminum sphere calculated with the standard method can approximate the ASF of calcification cluster in a background with anthropomorphic properties.

Daily, remotely controlled Quality Control (QC) in mammography has been a requirement of our breast cancer screening organization for accepting digital mammography in the screening programme. This set the scene for the development of an automated platform for daily QC supervision in digital mammography more than 15 years ago. While there was general acceptance of the need to supervise the first digital detectors installed in the mid-2000s, later digital systems have been more robust. One could question whether daily QC is of benefit for modern units. In this retrospective study we will illustrate and quantify the types of problems encountered with daily QC during the period January 2015 – January 2020. All issues were categorized as either localized artefacts (‘pixels’), line or line segment artefacts (‘lines’), special pattern or detector blocks visible (‘blocks’), lag and ghosts (‘ghosts’), sudden changes (in exposure) (‘reprod’), inhomogeneities, collimator problems and other. Trend analysis was then performed over the complete period. The number of issues found was 259, distributed over the categories as follows: 138 pixels, 28 lines, 14 blocks, 38 ghosts, 18 reprod, 15 inhomogeneities, 7 collimator problems and 1 other. Trend analysis revealed 44/103 stable systems, 54 systems with sudden changes in exposure (related to new software or new settings) and 7 of these 54 systems had periods with high variability in exposure settings. In 5 systems there was a steady but notable change in parameters over time. Daily QC applied to current units still detects a large number of issues. Remotely controlled daily QC supervision ensures constant quality of the systems on a daily basis and detects problems before they have any large impact on the clinical practice. The effort is considered worthwhile.

Geometric distortion is the inaccurate representation of the size or shape of a structure in the radiographic image. Exaggerated distortion makes radiography unacceptable for diagnosis. A new algorithm that was developed by us provides data on geometric distortion (GD) and ghost artifact-distortion (GAD) of digital breast tomosynthesis (DBT) images. This algorithm is similar to the one developed by the National Coordinating Centre for the Physics of Mammography (NCCPM), with the advantage of allowing the user to select the best-fit region of interest (ROI). The selection ensures that no information about the artifact dispersion contained in a ROI is lost. The aim of this study was to evaluate the dependence of ROI dimension (width and height) on the GD and GDA evaluation in digital breast tomosynthesis images using the new algorithm and to compare the results obtained with the limit values of reference, based on routine quality control tests for breast tomosynthesis. For the analyzes, the images were initially acquired with a 5 mm thick rectangular phantom composed of polymethyl methacrylate (PMMA) containing 1 mm diameter aluminum spheres. The phantom was inserted in the 60 mm thick PMMA phantom, positioned 25 mm away from the compression tray. The height of in-focus plane, the accuracy of positioning in the focus plane, and the appearance of aluminum spheres in the adjacent in-focus planes were analyzed for different ROI dimensions.

Publisher’s Note: This paper, originally published on 22 May 2020, was replaced with a corrected/revised version on 9 June 2020. If you downloaded the original PDF but are unable to access the revision, please contact SPIE Digital Library Customer Service for assistance.