Assessment of computed tomography coronary angiograms for diagnostic purposes is a mostly manual, timeconsuming task demanding a high degree of clinical experience. In order to support diagnosis, a method for reliable automatic detection of stenotic lesions in computed tomography angiograms is presented. Thereby, lesions are detected by boosting-based classification. Hence, a strong classifier is trained using the AdaBoost algorithm on annotated data. Subsequently, the resulting strong classification function is used in order to detect different types of coronary lesions in previously unseen data. As pattern recognition algorithms require a description of the objects to be classified, a novel approach for feature extraction in computed tomography angiograms is introduced. By generation of cylinder segments that approximate the vessel shape at multiple scales, feature values can be extracted that adequately describe the properties of stenotic lesions. As a result of the multi-scale approach, the algorithm is capable of dealing with the variability of stenotic lesion configuration. Evaluation of the algorithm was performed on a large database containing unseen segmented centerlines from cardiac computed tomography images. Results showed that the method was able to detect stenotic cardiovascular diseases with high sensitivity and specificity. Moreover, lesion based evaluation revealed that the majority of stenosis can be reliable identified in terms of position, type and extent.

We are investigating the feasibility of a computer-aided detection (CAD) system to assist radiologists in diagnosing coronary artery disease in ECG gated cardiac multi-detector CT scans having calcified plaque. Coronary artery stenosis analysis is challenging if calcified plaque or the iodinated blood pool hides viable lumen. The research described herein provides an improved presentation to the radiologist by removing obscuring calcified plaque and blood pool. The algorithm derives a Gaussian estimate of the point spread function (PSF) of the scanner responsible for plaque blooming by fitting measured CTA image profiles. An initial estimate of the extent of calcified plaque is obtained from the image evidence using a simple threshold. The Gaussian PSF estimate is then convolved with the initial plaque estimate to obtain an estimate of the extent of the blooming artifact and this plaque blooming image is subtracted from the CT image to obtain an image largely free of obscuring plaque. In a separate step, the obscuring blood pool is suppressed using morphological operations and adaptive region growing. After processing by our algorithm, we are able to project the segmented plaque-free lumen to form synthetic angiograms free from obstruction. We can also analyze the coronary arteries with vessel tracking and centerline extraction to produce cross sectional images for measuring lumen stenosis. As an additional aid to radiologists, we also produce plots of calcified plaque and lumen cross-sectional area along selected blood vessels. The method was validated using digital phantoms and actual patient data, including in one case, a validation against the results of a catheter angiogram.

Radiological bone age assessment is based on global or local image regions of interest (ROI), such as epiphyseal regions or the area of carpal bones. Usually, these regions are compared to a standardized reference and a score determining the skeletal maturity is calculated. For computer-assisted diagnosis, automatic ROI extraction is done so far by heuristic approaches. In this work, we apply a high-level approach of scene analysis for knowledge-based ROI segmentation. Based on a set of 100 reference images from the IRMA database, a so called structural prototype (SP) is trained. In this graph-based structure, the 14 phalanges and 5 metacarpal bones are represented by nodes, with associated location, shape, as well as texture parameters modeled by Gaussians. Accordingly, the Gaussians describing the relative positions, relative orientation, and other relative parameters between two nodes are associated to the edges. Thereafter, segmentation of a hand radiograph is done in several steps: (i) a multi-scale region merging scheme is applied to extract visually prominent regions; (ii) a graph/sub-graph matching to the SP robustly identifies a subset of the 19 bones; (iii) the SP is registered to the current image for complete scene-reconstruction (iv) the epiphyseal regions are extracted from the reconstructed scene. The evaluation is based on 137 images of Caucasian males from the USC hand atlas. Overall, an error rate of 32% is achieved, for the 6 middle distal and medial/distal epiphyses, 23% of all extractions need adjustments. On average 9.58 of the 14 epiphyseal regions were extracted successfully per image. This is promising for further use in content-based image retrieval (CBIR) and CBIR-based automatic bone age assessment.

A method to obtain 3D structural measurements of the proximal femur from 2D DXA images and a statistical atlas is presented. A statistical atlas of a proximal femur was created consisting of both 3D shape and volumetric density information and then deformably registered to 2D fan-beam DXA images. After the registration process, a series of 3D structural measurements were taken on QCT-estimates generated by transforming the registered statistical atlas into a voxel volume. These measurements were compared to the equivalent measurements taken on the actual QCT (ground truth) associated with the DXA images for each of 20 human cadaveric femora. The methodology and results are presented to address the potential clinical feasibility of obtaining 3D structural measurements from limited angle DXA scans and a statistical atlas of the proximal femur in-vivo.

The incidence of osteoporosis and associated fractures becomes an increasingly relevant issue for the public health institutions of industrialized nations. Fractures of the hip represent the worst complication of osteoporosis with a significantly elevated rate of mortality. Prediction of fracture risk is a major focus of osteoporosis research and, over the years, has been approched from different angles. There exist two distinct subtypes of transcervical and pertrochanteric hip fracture that can be distinguished on the basis of the anatomical location of the injury. While the epidemiology of hip fractures has been well described, typically, little or no distinction is made between the subtypes. The object of this study was to determine whether local topological texture properties based on the Minkowski Functionals (MF) obtained from standard radiographs of the proximal femur in patients with hip fracture can be used to differentiate between the two types of fracture pattern. The texture features were extracted from standardized regions of interest (femoral head, neck, and pertrochanteric region) in clinical radiographs of the hip obtained from 90 post-menopausal women (69.8 ± 7.9 yrs). 30 of the women had sustained pertrochanteric fractures, 30 had transcervical hip fractures and 30 were age-matched controls. We determined an optimized topological parameter MF2Dloc using an integrative filtering procedure based on a sliding-windows algorithm. Statistical relationship between the fracture type (pertrochanteric/transcervical) and the value of MF2Dloc was assessed by receiver-operator characteristic (ROC) analysis. Depending on the anatomical location of the region of interest for texture analysis correct classification of tanscervial and pertrochanteric fractures ranged from AUC = 0.79 to 0.98. In conclusion, quantitative texture properties of trabecular bone extracted from radiographs of the hip can be used to identify patients with hip fracture and to distinguish between pertrochanteric and transcervical fracture types. The degree of correct classification varies with choice of anatomical site for texture analysis. The results of our study may help to understand the mechanism of the two types of hip fracture.

Characterization and quantification of the severity of diffuse parenchymal lung diseases (DPLDs) using Computed Tomography (CT) is an important issue in clinical research. Recently, several classification-based computer-aided diagnosis (CAD) systems [1-3] for DPLD have been proposed. For some of those systems, a degradation of performance [2] was reported on unseen data because of considerable inter-patient variances of parenchymal tissue patterns. We believe that a CAD system of real clinical value should be robust to inter-patient variances and be able to classify unseen cases online more effectively. In this work, we have developed a novel adaptive knowledge-driven CT image search engine that combines offline learning aspects of classification-based CAD systems with online learning aspects of content-based image retrieval (CBIR) systems. Our system can seamlessly and adaptively fuse offline accumulated knowledge with online feedback, leading to an improved online performance in detecting DPLD in both accuracy and speed aspects. Our contribution lies in: (1) newly developed 3D texture-based and morphology-based features; (2) a multi-class offline feature selection method; and, (3) a novel image search engine framework for detecting DPLD. Very promising results have been obtained on a small test set.

Emphysema is a disease of the lungs that destroys the alveolar air sacs and induces long-term respiratory dysfunction. CT scans allow for imaging of the anatomical basis of emphysema and for visual assessment by radiologists of the extent present in the lungs. Several measures have been introduced for the quantification of the extent of disease directly from CT data in order to add to the qualitative assessments made by radiologists. In this paper we compare emphysema index, mean lung density, histogram percentiles, and the fractal dimension to visual grade in order to evaluate the predictability of radiologist visual scoring of emphysema from low-dose CT scans through quantitative scores, in order to determine which measures can be useful as surrogates for visual assessment. All measures were computed over nine divisions of the lung field (whole lung, individual lungs, and upper/middle/lower thirds of each lung) for each of 148 low-dose, whole lung scans. In addition, a visual grade of each section was also given by an expert radiologist. One-way ANOVA and multinomial logistic regression were used to determine the ability of the measures to predict visual grade from quantitative score. We found that all measures were able to distinguish between normal and severe grades (p<0.01), and between mild/moderate and all other grades (p<0.05). However, no measure was able to distinguish between mild and moderate cases. Approximately 65% prediction accuracy was achieved from using quantitative score to predict visual grade, with 73% if mild and moderate cases are considered as a single class.

Volumetric CT has the potential to improve the quantitative analysis of lung nodule size change compared to currently used one-dimensional measurement practices. Towards that goal, we have been conducting studies using an anthropomorphic phantom to quantify sources of volume measurement error. One source of error is the measurement technique or software tool used to estimate lesion volume. In this manuscript, we present a template-based approach which utilizes the properties of the acquisition and reconstruction system to quantify nodule volume. This approach may reduce the error associated with the volume estimation technique, thereby improving our ability to estimate the error directly associated with CT parameters and nodule characteristics. Our estimation approach consists of: (a) the simulation of the object-to-image transformation of a helical CT system, (b) the creation of a bank of simulated 3D nodule templates of varying sizes, and (c) the 3D matching of synthetic nodules - that were attached to lung vasculature and scanned with a 16-slice MDCT system - to the bank of simulated templates to estimate nodule volume. Results based on 10 repeat scans for different protocols and a root mean square error (RMSE) similarity metric showed a relative bias of 88%, 14%, and 4% for the measurement of 5 mm, 8 mm and 10 mm low density nodules (-630 HU) compared to -3%, -6%, and 8% for nodules of +100HU density. However, the relative bias for the small, low density nodules (5 mm, -630 HU), was significantly reduced to 7% when a penalized RMSE metric was used to enforce a symmetry constraint that reduced the impact of attached vessels. The results are promising for the use of this measurement approach as a low-bias estimator of nodule volume which will allow the systematic quantification and ranking of measurement error in volumetric CT analysis of lung nodules.

Documenting any change in airway dimensions over time may be relevant for monitoring the progression of pulmonary diseases. In order to correctly measure the change in segmental dimensions of airways, it is necessary to locate the identical airway segments across two scans. In this paper, we present an automated method to match individual bronchial segments from a pair of low-dose CT scans. Our method uses the intensity information in addition to the graph structure as evidences for matching the individual segments. 3D image correlation matching technique is employed to match the region of interest around the branch points in two scans and therefore locate the matching bronchial segments. The matching process was designed to address the differences in airway tree structures from two scans due to the variation in tree segmentations. The algorithm was evaluated using 114 pairs of low-dose CT scans (120 kV, 40 mAs). The total number of segments matched was 3591, of which 99.7% were correctly matched. When the matching was limited to the bronchial segments of the fourth generation or less, the algorithm correctly identified all of 1553 matched segments.

Computed tomographic pulmonary angiography (CTPA) has been reported to be an effective means for clinical diagnosis of pulmonary embolism (PE). We are developing a computer-aided diagnosis (CAD) system for assisting radiologists in detection of pulmonary embolism in CTPA images. The pulmonary vessel tree is extracted based on the analysis of eigenvalues of Hessian matrices at multiple scales followed by 3D hierarchical EM segmentation. A multiprescreening method is designed to identify suspicious PEs along the extracted vessels. A linear discriminant analysis (LDA) classifier with feature selection is then used to reduce false positives (FPs). Two data sets of 59 and 69 CTPA PE cases were randomly selected from patient files at the University of Michigan (UM) and the PIOPED II study, respectively, and used as independent training and test sets. The PEs that were identified by three experienced thoracic radiologists were used as the gold standard. The detection performance of the CAD system was assessed by free response receiver operating characteristic analysis. The results indicated that our PE detection system can achieve a sensitivity of 80% at 18.9 FPs/case on the PIOPED cases when the LDA classifier was trained with the UM cases. The test sensitivity with the UM cases is 80% at 22.6 FPs/cases when the LDA classifier was trained with the PIOPED cases.

Many automated techniques have been proposed to classify diffuse lung disease patterns. Most of the techniques utilize texture analysis approaches with second and higher order statistics, and show successful classification result among various lung tissue patterns. However, the approaches do not work well for the patterns with inhomogeneous texture distribution within a region of interest (ROI), such as reticular and honeycombing patterns, because the statistics can only capture averaged feature over the ROI. In this work, we have introduced the bag-of-features approach to overcome this difficulty. In the approach, texture images are represented as histograms or distributions of a few basic primitives, which are obtained by clustering local image features. The intensity descriptor and the Scale Invariant Feature Transformation (SIFT) descriptor are utilized to extract the local features, which have significant discriminatory power due to their specificity to a particular image class. In contrast, the drawback of the local features is lack of invariance under translation and rotation. We improved the invariance by sampling many local regions so that the distribution of the local features is unchanged. We evaluated the performance of our system in the classification task with 5 image classes (ground glass, reticular, honeycombing, emphysema, and normal) using 1109 ROIs from 211 patients. Our system achieved high classification accuracy of 92.8%, which is superior to that of the conventional system with the gray level co-occurrence matrix (GLCM) feature especially for inhomogeneous texture patterns.

It is important for diagnosis of pulmonary diseases to measure volume of accumulating pleural effusion in threedimensional thoracic CT images quantitatively. However, automated extraction of pulmonary effusion correctly is difficult. Conventional extraction algorithm using a gray-level based threshold can not extract pleural effusion from thoracic wall or mediastinum correctly, because density of pleural effusion in CT images is similar to those of thoracic wall or mediastinum. So, we have developed an automated extraction method of pulmonary effusion by use of extracting lung area with pleural effusion. Our method used a template of lung obtained from a normal lung for segmentation of lungs with pleural effusions. Registration process consisted of two steps. First step was a global matching processing between normal and abnormal lungs of organs such as bronchi, bones (ribs, sternum and vertebrae) and upper surfaces of livers which were extracted using a region-growing algorithm. Second step was a local matching processing between normal and abnormal lungs which were deformed by the parameter obtained from the global matching processing. Finally, we segmented a lung with pleural effusion by use of the template which was deformed by two parameters obtained from the global matching processing and the local matching processing. We compared our method with a conventional extraction method using a gray-level based threshold and two published methods. The extraction rates of pleural effusions obtained from our method were much higher than those obtained from other methods. Automated extraction method of pulmonary effusion by use of extracting lung area with pleural effusion is promising for diagnosis of pulmonary diseases by providing quantitative volume of accumulating pleural effusion.

An estimation of the so called Ground Truth (GT), i.e. the actual lesion region, can minimize readers' subjectivity if multiple readers' markings are combined. Two methods perform this estimate by considering the spatial location of voxels: Thresholded Probability-Map (TPM) and Simultaneous Truth and Performance Level Estimation (STAPLE). An analysis of these two methods has already been performed. The purpose of this study, however, is gaining a new insight into the method outcomes by comparing the estimated regions. A subset of the publicly available Lung Image Database Consortium archive was used, selecting pulmonary nodules documented by all four radiologists. The TPM estimator was computed by assigning to each voxel a value equal to average number of readers that included such voxel in their markings and then applying a threshold of 0.5. Our STAPLE implementation is loosely based on a version from ITK, to which we added the graph cut post-processing. The pair-wise similarities between the estimated ground truths were analyzed by computing the respective Jaccard coefficients. Then, the sign test of the differences between the volumes of TPM and STAPLE was performed. A total of 35 nodules documented on 26 scans by all four radiologists were available. The spatial agreement had a one-sided 90% Confidence Interval of [0.92, 1.00]. The sign test of the differences had a p-value less than 0.001. We found that (a) the differences in their volume estimates are statistically significant, (b) the spatial disagreement between the two estimators is almost completely due to the exclusion of voxels marked by exactly two readers, (c) STAPLE tends to weight more, in its GT estimate, readers marking broader regions.

In case-based computer-aided decision systems (CB-CAD) a query case is compared to known examples stored in the systems case base (also called a reference library). These systems offer competitive classification performance and are easy to expand. However, they also require efficient management of the case base. As CB-CAD systems are becoming more popular, the problem of case base optimization has recently attracted interest among CAD researchers. In this paper we present preliminary results of a study comparing several case base reduction techniques. We implemented six techniques previously proposed in machine learning literature and applied it to the classification problem of distinguishing masses and normal tissue in mammographic regions of interest. The results show that the random mutation hill climbing technique offers a drastic reduction of the number of case base examples while providing a significant improvement in classification performance. Random selection allowed for reduction of the case base to 30% without notable decrease in performance. The remaining techniques (i.e., condensed nearest neighbor, reduced nearest neighbor, edited nearest neighbor, and All k-NN) resulted in moderate reduction (to 50-70% of the original size) at the cost of decrease in CB-CAD performance.

Purpose: To develop, test, and evaluate a novel unsupervised machine learning method for computer-aided diagnosis and analysis of multidimensional data, such as biomedical imaging data. Methods: We introduce the Exploration Machine (XOM) as a method for computing low-dimensional representations of high-dimensional observations. XOM systematically inverts functional and structural components of topology-preserving mappings. By this trick, it can contribute to both structure-preserving visualization and data clustering. We applied XOM to the analysis of whole-genome microarray imaging data, comprising 2467 79-dimensional gene expression profiles of Saccharomyces cerevisiae, and to model-free analysis of functional brain MRI data by unsupervised clustering. For both applications, we performed quantitative comparisons to results obtained by established algorithms. Results: Genome data: Absolute (relative) Sammon error values were 5.91·10⁵ (1.00) for XOM, 6.50·10⁵ (1.10) for Sammon's mapping, 6.56·10⁵ (1.11) for PCA, and 7.24·105 (1.22) for Self-Organizing Map (SOM). Computation times were 72, 216, 2, and 881 seconds for XOM, Sammon, PCA, and SOM, respectively. - Functional MRI data: Areas under ROC curves for detection of task-related brain activation were 0.984 ± 0.03 for XOM, 0.983 ± 0.02 for Minimal-Free-Energy VQ, and 0.979 ± 0.02 for SOM. Conclusion: For both multidimensional imaging applications, i.e. gene expression visualization and functional MRI clustering, XOM yields competitive results when compared to established algorithms. Its surprising versatility to simultaneously contribute to dimensionality reduction and data clustering qualifies XOM to serve as a useful novel method for the analysis of multidimensional data, such as biomedical image data in computer-aided diagnosis.

We have analyzed 3000 nodule delineations and malignancy ratings of pulmonary nodules made by expert observers in the IDRI CT lung image database. The agreement between nodule volume from automatic segmentation and expert delineations is almost as high as inter-observer agreement. For the experts' malignancy rating the inter-observer agreement is quite modest. Linear and support vector regression models have been tested to emulate the expert malignancy rating from a small number of automatically computed numerical image features. Machine-observer and inter-observer agreement have been evaluated using linear correlation and weighted kappa coefficient. The results suggest that the numerical computed malignancy - if considered as an additional observer - cannot be distinguished from the expert ratings.

Automated image analysis is used for risk assessment and computer aided diagnosis of breast cancer. A prerequisite for this automation is an efficient and robust segmentation of the region of interest (ROI). Extraction of breast without the pectoral muscle, the ROI in our study, from mammograms is a challenging task due to the tapering nature of the breast at the skin-air interface and the overlap between the high density regions and the pectoral muscle in the medio-lateral oblique (MLO) views. To segment breast skin-air interface, Otsu's multilevel threshold based algorithm constrained with a shape prior is used. Starting at a coarse scale, the pectoral muscle is detected by fitting a novel adaptive statistical shape model at the top left corner in the medio-lateral oblique views of mammograms and its position is localized by scale focusing. A novel energy term is proposed to fit the shape model. The proposed algorithm is applied to a set of 50 mammograms as a post hoc analysis of an earlier trial. The results are evaluated by comparing with manual annotations. In the pectoral muscle detection the metrics used for the evaluation are: (a) mean Hausdorff distance (MHD):6.7 pixels; (b) area overlap:87%; (c) false positive rate (FP):0.7%; (d) false negative rate (FN):12.1% and (e) mean relative breast area error (MRBAE):101.19±5.26 %.

The purpose of this study is to investigate three-class Bayesian artificial neural networks (BANN) in dynamic contrastenhanced MRI (DCE-MRI) CAD in distinguishing different types of breast lesions including ductal carcinoma in situ (DCIS), invasive ductal carcinoma (IDC), and benign. The database contains 72 DCIS lesions, 124 IDC lesions, and 131 benign breast lesions (no cysts). Breast MR images were obtained with a clinical DCE-MRI scanning protocol. In 3D, we automatically segmented each lesion and calculated its characteristic kinetic curve using the fuzzy c-means method. Morphological and kinetic features were automatically extracted, and stepwise linear discriminant analysis was utilized for feature selection in four subcategories: DCIS vs. IDC, DCIS vs. benign, IDC vs. benign, and malignant (DCIS + IDC) vs. benign. Classification was automatically performed with the selected features for each subcategory using round-robin-by-lesion two-class BANN and three-class BANN. The performances of the classifiers were assessed with two-class ROC analysis. We failed to show any statistically significant differences between the two-class BANN and three-class BANN for all four classification tasks, demonstrating that the three-class BANN performed similarly to the two-class BANN. A three-class BANN is expected to be more desirable in the clinical arena for both diagnosis and patient management.

Breast cancer diagnosis based on magnetic resonance images (breast MRI) is increasingly being accepted as an additional diagnostic tool to mammography and ultrasound, with distinct clinical indications.¹ Its capability to detect and differentiate lesion types with high sensitivity and specificity is countered by the fact that visual human assessment of breast MRI requires long experience. Moreover, the lack of evaluation standards causes diagnostic results to vary even among experts. The most important MR acquisition technique is dynamic contrast enhanced (DCE) MR imaging since different lesion types accumulate contrast material (CM) differently. The wash-in and wash-out characteristic as well as the morphologic characteristic recorded and assessed from MR images therefore allows to differentiate benign from malignant lesions. In this work, we propose to calculate second order statistical features (Haralick textures) for given lesions based on subtraction and 4D images and on parametermaps. The lesions are classified with a linear classification scheme into probably malignant or probably benign. The method and model was developed on 104 histologically graded lesions (69 malignant and 35 benign). The area under the ROC curve obtained is 0.91 and is already comparable to the performance of a trained radiologist.

A computer-aided diagnosis (CADx) system with the ability to predict the probability of malignancy (PM) of a mass can potentially assist radiologists in making correct diagnostic decisions. In this study, we designed a CADx system using logistic regression (LR) as the feature classifier which could estimate the PM of a mass. Our data set included 488 ultrasound (US) images from 250 biopsy-proven breast masses (100 malignant and 150 benign). The data set was divided into two subsets T1 and T2. Two experienced radiologists, R1 and R2, independently provided Breast Imaging Reporting and Data System (BI-RADS) assessments and PM ratings for data subsets T2 and T1, respectively. An LR classifier was designed to estimate the PM of a mass using two-fold cross validation, in which the data subsets T1 and T2 served once as the training and once as the test set. To evaluate the performance of the system, we compared the PM estimated by the CADx system with radiologists' PM ratings (12-point scale) and BI-RADS assessments (6-point scale). The correlation coefficients between the PM ratings estimated by the radiologists and by the CADx system were 0.71 and 0.72 for data subsets T1 and T2, respectively. For the BI-RADS assessments provided by the radiologists and estimated by the CADx system, the correlation coefficients were 0.60 and 0.67 for data subsets T1 and T2, respectively. Our results indicate that the CADx system may be able to provide not only a malignancy score, but also a more quantitative estimate for the PM of a breast mass.

For the detection of microcalcifications, accurate noise estimation has shown to be an important step. In tomosynthesis, noise models have been proposed for projection data. However, it is expected that manufacturers of tomosynthesis systems will not store the raw projection images, but only the reconstructed volumes. We therefore investigated if and how signal dependent image noise can be modelled in the reconstructed volumes. For this research we used a dataset of 41 tomosynthesis volumes, of which 12 volumes contained a total of 20 microcalcification clusters. All volumes were acquired with a prototype of Sectra's photon-counting tomosynthesis system. Preliminary results show that image noise is signal dependent in a reconstructed volume, and that a model of this noise can be estimated from a volume at hand. Evaluation of the noise model was performed by using a basic microcalcification cluster detection algorithm that classifies voxels by using a threshold on a local contrast filter. Image noise was normalized by dividing local contrast in a voxel by the standard deviation of the estimated image noise in that voxel. FROC analysis shows that performance increases strongly, when we use our model to correct for signal dependent image noise in reconstructed volumes.

Breast density has been shown to be associated with the risk of developing breast cancer, and MRI has been recommended for high-risk women screening, however, it is still unknown how the breast parenchymal enhancement on DCE-MRI is associated with breast density and breast cancer risk. Ninety-two DCE-MRI exams of asymptomatic women with normal MR findings were included in this study. The 3D breast volume was automatically segmented using a volume-growing based algorithm. The extracted breast volume was classified into fibroglandular and fatty regions based on the discriminant analysis method. The parenchymal kinetic curves within the breast fibroglandular region were extracted and categorized by use of fuzzy c-means clustering, and various parenchymal kinetic characteristics were extracted from the most enhancing voxels. Correlation analysis between the computer-extracted percent dense measures and radiologist-noted BIRADS density ratings yielded a correlation coefficient of 0.76 (p<0.0001). From kinetic analyses, 70% (64/92) of most enhancing curves showed persistent curve type and reached peak parenchymal intensity at the last postcontrast time point; with 89% (82/92) of most enhancing curves reaching peak intensity at either 4th or 5th post-contrast time points. Women with dense breast (BIRADS 3 and 4) were found to have more parenchymal enhancement at their peak time point (Ep) with an average Ep of 116.5% while those women with fatty breasts (BIRADS 1 and 2) demonstrated an average Ep of 62.0%. In conclusion, breast parenchymal enhancement may be associated with breast density and may be potential useful as an additional characteristic for assessing breast cancer risk.

Since different imaging modalities provide complementary information regarding the same lesion, combining information from different modalities may increase diagnostic accuracy. In this study, we investigated the use of computerized features of lesions imaged via both full-field digital mammography (FFDM) and dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) in the classification of breast lesions. Using a manually identified lesion location, i.e. a seed point on FFDM images or a ROI on DCE-MRI images, the computer automatically segmented mass lesions and extracted a set of features for each lesion. Linear stepwise feature selection was firstly performed on single modality, yielding one feature subset for each modality. Then, these selected features served as the input to another feature selection procedure when extracting useful information from both modalities. The selected features were merged by linear discriminant analysis (LDA) into a discriminant score. Receiver operating characteristic (ROC) analysis was used to evaluate the performance of the selected feature subset in the task of distinguishing between malignant and benign lesions. From a FFDM database with 321 lesions (167 malignant and 154 benign), and a DCE-MRI database including 181 lesions (97 malignant and 84 benign), we constructed a multi-modality dataset with 51 lesions (29 malignant and 22 benign). With leave-one-out-by-lesion evaluation on the multi-modality dataset, the mammography-only features yielded an area under the ROC curve (AUC) of 0.62 ± 0.08 and the DCE-MRI-only features yielded an AUC of 0.94±0.03. The combination of these two modalities, which included a spiculation feature from mammography and a kinetic feature from DCE-MRI, yielded an AUC of 0.94. The improvement of combining multi-modality information was statistically significant as compared to the use of mammography only (p=0.0001). However, we failed to show the statistically significant improvement as compared to DCE-MRI, with the limited multi-modality dataset (p=0.22).

We propose a new method to detect architectural distortions and spiculated masses in digital breast tomosynthesis volumes. To achieve this goal, an a contrario approach is used. In this approach, an event, corresponding to a minimal number of structures converging toward the same location, is defined such that its expectation of occurrence within a random image is very low. Occurrences of this event in real images are then detected and considered as possible lesion locations. During the last step, the number of false positives is reduced through classification using attributes computed on histograms of structure orientations. The approach was tested using the leave-one-out method on a database composed of 38 breasts (10 containing a lesion and 28 containing no lesion). A sensitivity of 0.8 at 1.68 false positives/breast was achieved.

In breast cancer screening, radiologists not only look at local properties of suspicious regions in the mammogram but take also into account more general contextual information. In this study we investigated the use of similar information for computer aided detection of malignant masses. We developed a new set of features that combine information from the candidate mass region and the whole image or mammogram. The developed context features were constructed to give information about suspiciousness of a region relative to other areas in the mammogram, the location in the image, the location in relation to dense tissue and the overall amount of dense tissue in the mammogram. We used a step-wise floating feature selection algorithm to select subsets from the set of available features. Feature selection was performed two times, once using the complete feature set (37 context and 40 local features) and once using only the local features. It was found that in the subsets selected from the complete feature set 30-60% were context features. At most one local feature present in the subset containing context features was not present in the subset without context features. We validated the performance of the selected subsets on a separate data set using cross validation and bootstrapping. For each subset size we compared the performance obtained using the features selected from the complete feature set to the performance obtained using the features selected from the local feature set. We found that subsets containing context features performed significantly better than feature sets containing no context features.

We have been developing a computerized scheme for selecting visually similar images that would be useful to radiologists in the diagnosis of masses on mammograms. Based on the results of the observer performance study, the presentation of similar images was useful, especially for less experienced observers. The test cases included 50 benign and 50 malignant masses. Ten observers, including five breast radiologists and five residents, were asked to provide the confidence level of the lesions being malignant before and after the presentation of similar images. By use of multireader, multi-case receiver operating characteristic analysis, the average areas under the curves for the five residents were 0.880 and 0.896 without and with similar images, respectively (p=0.040). There were four malignant cases in which the initial ratings were relatively low, but the similar images alerted the residents to increase their confidence levels of malignancy close to those by the breast radiologists. The presentation of similar images may cause some observers falsely to increase their suspicion for some benign cases; however, if similar images can alert radiologists to recognize the signs of malignancy and also help them to decrease their suspicion correctly for some benign cases, they can be useful in the diagnosis on mammograms.

We are developing a CAD system for mass detection on digital breast tomosynthesis (DBT) mammograms. In this study, we compared the detection accuracy on DBT and conventional screen-film mammograms (SFMs). DBT mammograms were acquired with a GE prototype system at the University of Michigan. 47 cases containing the CC- and MLO-view DBT mammograms of the breast with a biopsy-proven mass and the corresponding two-view SFMs of the same breast were collected. Subjective judgment showed that the masses were much more conspicuous on DBT slices than on SFMs. The CAD system for DBT includes two parallel processes, one performs mass detection in the reconstructed DBT volume, and the other in the projection view (PV) images. The mass likelihood scores estimated for each mass candidate in the two processes are merged to differentiate masses and false positives (FPs). For detection on SFMs, we previously developed a dual system approach by fusing two single CAD systems optimized for detection of average and subtle masses, respectively. A trained neural network is used to merge the mass likelihood scores of the two single systems to reduce FPs. At the case-based sensitivities of 80% and 85%, mass detection in the DBT volume resulted in an average of 0.72 and 1.06 FPs/view, and detection in the SFMs yielded 0.94 and 1.67 FPs/view, respectively. The difference fell short of statistical significance (p=0.07) by JAFROC analysis. Study is underway to collect a larger data set and to further improve the DBT CAD system.

With the introduction of Full Field Digital Mammography (FFDM) accurate automatic volumetric breast density (VBD) estimation has become possible. As VBD enables the design of features that incorporate 3D properties, these methods offer opportunities for computer aided detection schemes. In this study we use VBD to develop features that represent how well a segmented region resembles the projection of a spherical object. The idea behind this is that due to compression of the breast, glandular tissue is likely to be compressed to a disc like shape, whereas cancerous tissue, being more difficult to compress, will retain its uncompressed shape. For each pixel in a segmented region we calculate the predicted dense tissue thickness assuming that the lesion has a spherical shape. The predicted thickness is then compared to the observed thickness by calculating the slope of a linear function relating the two. In addition we calculate the variance of the error of the fit. To evaluate the contribution of the developed VBD features to our CAD system we use an FFDM dataset consisting of 266 cases, of which 103 were biopsy proven malignant masses and 163 normals. It was found that compared to the false positives, a large fraction of the true positives has a slope close to 1.0 indicating that the true positives fit the modeled spheres best. When the VBD based features were added to our CAD system, aimed at the detection and classification of malignant masses, a small but significant increase in performance was achieved.

Lung cancer is the leading cause of cancer death in the United States. In lung-cancer staging, central-chest lymph nodes and associated nodal stations, as observed in three-dimensional (3D) multidetector CT (MDCT) scans, play a vital role. However, little work has been done in relation to lymph nodes, based on MDCT data, due to the complicated phenomena that give rise to them. Using our custom computer-based system for 3D MDCT-based pulmonary lymph-node analysis, we conduct a detailed study of lymph nodes as depicted in 3D MDCT scans. In this work, the Mountain lymph-node stations are automatically defined by the system. These defined stations, in conjunction with our system's image processing and visualization tools, facilitate lymph-node detection, classification, and segmentation. An expert pulmonologist, chest radiologist, and trained technician verified the accuracy of the automatically defined stations and indicated observable lymph nodes. Next, using semi-automatic tools in our system, we defined all indicated nodes. Finally, we performed a global quantitative analysis of the characteristics of the observed nodes and stations. This study drew upon a database of 32 human MDCT chest scans. 320 Mountain-based stations (10 per scan) and 852 pulmonary lymph nodes were defined overall from this database. Based on the numerical results, over 90% of the automatically defined stations were deemed accurate. This paper also presents a detailed summary of central-chest lymph-node characteristics for the first time.

Computed tomography (CT) of the chest is a very common staging investigation for the assessment of mediastinal, hilar, and intrapulmonary lymph nodes in the context of lung cancer. In the current clinical workflow, the detection and assessment of lymph nodes is usually performed manually, which can be error-prone and timeconsuming. We therefore propose a method for the automatic detection of mediastinal, hilar, and intrapulmonary lymph node candidates in contrast-enhanced chest CT. Based on the segmentation of important mediastinal anatomy (bronchial tree, aortic arch) and making use of anatomical knowledge, we utilize Hessian eigenvalues to detect lymph node candidates. As lymph nodes can be characterized as blob-like structures of varying size and shape within a specific intensity interval, we can utilize these characteristics to reduce the number of false positive candidates significantly. We applied our method to 5 cases suspected to have lung cancer. The processing time of our algorithm did not exceed 6 minutes, and we achieved an average sensitivity of 82.1% and an average precision of 13.3%.

Early response assessment of cancer therapy is a crucial component towards a more effective and patient individualized cancer therapy. Integrated PET/CT systems provide the opportunity to combine morphologic with functional information. We have developed algorithms which allow the user to track both tumor volume and standardized uptake value (SUV) measurements during the therapy from series of CT and PET images, respectively. To prepare for tumor volume estimation we have developed a new technique for a fast, flexible, and intuitive 3D definition of meshes. This initial surface is then automatically adapted by means of a model-based segmentation algorithm and propagated to each follow-up scan. If necessary, manual corrections can be added by the user. To determine SUV measurements a prioritized region growing algorithm is employed. For an improved workflow all algorithms are embedded in a PET/CT therapy monitoring software suite giving the clinician a unified and immediate access to all data sets. Whenever the user clicks on a tumor in a base-line scan, the courses of segmented tumor volumes and SUV measurements are automatically identified and displayed to the user as a graph plot. According to each course, the therapy progress can be classified as complete or partial response or as progressive or stable disease. We have tested our methods with series of PET/CT data from 9 lung cancer patients acquired at Princess Margaret Hospital in Toronto. Each patient underwent three PET/CT scans during a radiation therapy. Our results indicate that a combination of mean metabolic activity in the tumor with the PET-based tumor volume can lead to an earlier response detection than a purely volume based (CT diameter) or purely functional based (e.g. SUV max or SUV mean) response measures. The new software seems applicable for easy, faster, and reproducible quantification to routinely monitor tumor therapy.

A new emphysema treatment uses endobronchial valves to perform lobar volume reduction. The degree of fissure completeness may predict treatment efficacy. This study investigated the behavior of a semiautomated algorithm for quantifying lung fissure integrity in CT with respect to reconstruction kernel and dose. Raw CT data was obtained for six asymptomatic patients from a high-risk population for lung cancer. The patients were scanned on either a Siemens Sensation 16 or 64, using a low-dose protocol of 120 kVp, 25 mAs. Images were reconstructed using kernels ranging from smooth to sharp (B10f, B30f, B50f, B70f). Research software was used to simulate an even lower-dose acquisition of 15 mAs, and images were generated at the same kernels resulting in 8 series per patient. The left major fissure was manually contoured axially at regular intervals, yielding 37 contours across all patients. These contours were read into an image analysis and pattern classification system which computed a Fissure Integrity Score (FIS) for each kernel and dose. FIS values were analyzed using a mixed-effects model with kernel and dose as fixed effects and patient as random effect to test for difference due to kernel and dose. Analysis revealed no difference in FIS between the smooth kernels (B10f, B30f) nor between sharp kernels (B50f, B70f), but there was a significant difference between the sharp and smooth groups (p = 0.020). There was no significant difference in FIS between the two low-dose reconstructions (p = 0.882). Using a cutoff of 90%, the number of incomplete fissures increased from 5 to 10 when the imaging protocol changed from B50f to B30f. Reconstruction kernel has a significant effect on quantification of fissure integrity in CT. This has potential implications when selecting patients for endobronchial valve therapy.

Features calculated from different dimensions of images capture quantitative information of the lung nodules through one or multiple image slices. Previously published computer-aided diagnosis (CADx) systems have used either twodimensional (2D) or three-dimensional (3D) features, though there has been little systematic analysis of the relevance of the different dimensions and of the impact of combining different dimensions. The aim of this study is to determine the importance of combining features calculated in different dimensions. We have performed CADx experiments on 125 pulmonary nodules imaged using multi-detector row CT (MDCT). The CADx system computed 192 2D, 2.5D, and 3D image features of the lesions. Leave-one-out experiments were performed using five different combinations of features from different dimensions: 2D, 3D, 2.5D, 2D+3D, and 2D+3D+2.5D. The experiments were performed ten times for each group. Accuracy, sensitivity and specificity were used to evaluate the performance. Wilcoxon signed-rank tests were applied to compare the classification results from these five different combinations of features. Our results showed that 3D image features generate the best result compared with other combinations of features. This suggests one approach to potentially reducing the dimensionality of the CADx data space and the computational complexity of the system while maintaining diagnostic accuracy.

This work involves the computer-aided diagnosis (CAD) of pulmonary embolism (PE) in contrast-enhanced computed tomography pulmonary angiography (CTPA). Contrast plays an important role in analyzing and identifying PE in CTPA. At times the contrast mixing in blood may be insufficient due to several factors such as scanning speed, body weight and injection duration. This results in a suboptimal study (mixing artifact) due to non-homogeneous enhancement of blood's opacity. Most current CAD systems are not optimized to detect PE in sub optimal studies. To this effect, we propose new techniques for CAD to work robustly in both optimal and suboptimal situations. First, the contrast level at the pulmonary trunk is automatically detected using a landmark detection tool. This information is then used to dynamically configure the candidate generation (CG) and classification stages of the algorithm. In CG, a fast method based on tobogganing is proposed which also detects wall-adhering emboli. In addition, our proposed method correctly encapsulates potential PE candidates that enable accurate feature calculation over the entire PE candidate. Finally a classifier gating scheme has been designed that automatically switches the appropriate classifier for suboptimal and optimal studies. The system performance has been validated on 86 real-world cases collected from different clinical sites. Results show around 5% improvement in the detection of segmental PE and 6% improvement in lobar and sub segmental PE with a 40% decrease in the average false positive rate when compared to a similar system without contrast detection.

In this paper we propose a technique for automatic detection of intracranial hemorrhage (ICH) and acute intracranial hemorrhage (AIH) in brain Computed Tomography (CT) for trauma cases where no contrast can be applied and the CT has large slice thickness. ICH or AIH comprise of internal bleeding (intra-axial) or external (extra-axial) to the brain substance. Large bleeds like in intra-axial region are easy to diagnose whereas it can be challenging if small bleed occurs in extra-axial region particularly in the absence of contrast. Bleed region needs to be distinguished from bleed-look-alike brain regions which are abnormally bright falx and fresh flowing blood. We propose an algorithm for detection of brain bleed in various anatomical locations. A preprocessing step is performed to segment intracranial contents and enhancement of region of interests(ROIs). A number of bleed and bleed-look-alike candidates are identified from a set of 11 available cases. For each candidate texture based features are extracted from non-separable quincunx wavelet transform along with some other descriptive features. The candidates are randomly divided into a training and test set consisting of both bleed and bleed-look- alike. A supervised classifier is designed based on the training sample features. A performance accuracy of 96% is attained for the independent test candidates.

FDG-PET is increasingly used for the evaluation of dementia patients, as major neurodegenerative disorders, such as Alzheimer's disease (AD), Lewy body dementia (LBD), and Frontotemporal dementia (FTD), have been shown to induce specific patterns of regional hypo-metabolism. However, the interpretation of FDG-PET images of patients with suspected dementia is not straightforward, since patients are imaged at different stages of progression of neurodegenerative disease, and the indications of reduced metabolism due to neurodegenerative disease appear slowly over time. Furthermore, different diseases can cause rather similar patterns of hypo-metabolism. Therefore, classification of FDG-PET images of patients with suspected dementia may lead to misdiagnosis. This work aims to find an optimal subset of features for automated classification, in order to improve classification accuracy of FDG-PET images in patients with suspected dementia. A novel feature selection method is proposed, and performance is compared to existing methods. The proposed approach adopts a combination of balanced class distributions and feature selection methods. This is demonstrated to provide high classification accuracy for classification of FDG-PET brain images of normal controls and dementia patients, comparable with alternative approaches, and provides a compact set of features selected.

We propose a single, quantitative metric called the disease evaluation factor (DEF) and assess its efficiency at estimating disease burden in normal, control subjects (CTRL) and probable Alzheimer's disease (AD) patients. The study group consisted in 75 patients with a diagnosis of probable AD and 75 age-matched normal CTRL without neurological or neuropsychological deficit. We calculated a reference eigenspace of MRI appearance from reference data, in which our CTRL and probable AD subjects were projected. We then calculated the multi-dimensional hyperplane separating the CTRL and probable AD groups. The DEF was estimated via a multidimensional weighted distance of eigencoordinates for a given subject and the CTRL group mean, along salient principal components forming the separating hyperplane. We used quantile plots, Kolmogorov-Smirnov and χ² tests to compare the DEF values and test that their distribution was normal. We used a linear discriminant test to separate CTRL from probable AD based on the DEF factor, and reached an accuracy of 87%. A quantitative biomarker in AD would act as an important surrogate marker of disease status and progression.

Breast Cancer is one of the most deadly cancers affecting middle-aged women. Accurate diagnosis and prognosis are crucial to reduce the high death rate. Nowadays there are numerous diagnostic tools for breast cancer diagnosis. In this paper we discuss a role of nuclear segmentation from fine needle aspiration biopsy (FNA) slides and its influence on malignancy classification. Classification of malignancy plays a very important role during the diagnosis process of breast cancer. Out of all cancer diagnostic tools, FNA slides provide the most valuable information about the cancer malignancy grade which helps to choose an appropriate treatment. This process involves assessing numerous nuclear features and therefore precise segmentation of nuclei is very important. In this work we compare three powerful segmentation approaches and test their impact on the classification of breast cancer malignancy. The studied approaches involve level set segmentation, fuzzy c-means segmentation and textural segmentation based on co-occurrence matrix. Segmented nuclei were used to extract nuclear features for malignancy classification. For classification purposes four different classifiers were trained and tested with previously extracted features. The compared classifiers are Multilayer Perceptron (MLP), Self-Organizing Maps (SOM), Principal Component-based Neural Network (PCA) and Support Vector Machines (SVM). The presented results show that level set segmentation yields the best results over the three compared approaches and leads to a good feature extraction with a lowest average error rate of 6.51% over four different classifiers. The best performance was recorded for multilayer perceptron with an error rate of 3.07% using fuzzy c-means segmentation.

Morphological and architectural characteristics of primary tissue compartments, such as epithelial nuclei (EN) and cytoplasm, provide important cues for cancer diagnosis, prognosis, and therapeutic response prediction. We propose two feature sets for the robust quantification of these characteristics in multiplex immunofluorescence (IF) microscopy images of prostate biopsy specimens. To enable feature extraction, EN and cytoplasm regions were first segmented from the IF images. Then, feature sets consisting of the characteristics of the minimum spanning tree (MST) connecting the EN and the fractal dimension (FD) of gland boundaries were obtained from the segmented compartments. We demonstrated the utility of the proposed features in prostate cancer recurrence prediction on a multi-institution cohort of 1027 patients. Univariate analysis revealed that both FD and one of the MST features were highly effective for predicting cancer recurrence (p ≤ 0.0001). In multivariate analysis, an MST feature was selected for a model incorporating clinical and image features. The model achieved a concordance index (CI) of 0.73 on the validation set, which was significantly higher than the CI of 0.69 for the standard multivariate model based solely on clinical features currently used in clinical practice (p < 0.0001). The contributions of this work are twofold. First, it is the first demonstration of the utility of the proposed features in morphometric analysis of IF images. Second, this is the largest scale study of the efficacy and robustness of the proposed features in prostate cancer prognosis.

High resolution digital pathology images have a wide range of variability in color, shape, size, number, appearance, location, and texture. The segmentation problem is challenging in this environment. We introduce a hybrid method that combines parametric machine learning with heuristic methods for feature extraction as well as pre- and post-processing steps for localizing diverse tissues in slide images. The method uses features such as color, intensity, texture, and spatial distribution. We use principal component analysis for feature reduction and train a two layer back propagation neural network (with one hidden layer). We perform image labeling at pixel-level and achieve higher than 96% automatic localization accuracy on 294 test images.

We are presenting continuing development of an information-theoretic (IT) CADe system for location-specific interrogation of screening mammograms to detect masses. IT-CADe relies on a knowledge library of mammographic cases with known ground truth and an evidence-based approach to make a decision regarding a query case. If the query is more similar to abnormal cases stored in the library, then the query is deemed also abnormal. Case similarity is measured using mutual information (MI). MI takes into account only the probabilities of the underlying image pixels but not their relative significance in the image. To address this limitation, we investigated a novel modification of the MI similarity measure by incorporating the saliency of image pixels. Specifically, a Gaussian saliency map was applied where central image pixels were given a higher weight and pixels' importance degraded progressively towards the image periphery. This map makes intuitively sense. If a mass is suspected at a particular location, then image pixels surrounding this location should be given higher importance in the MI calculation than pixels further away from this specific location. The new MI measure was tested with a leave-one-out scheme on a database of 1,820 mammographic regions (901 with masses and 919 normal). Further validation was performed on additional datasets of mammographic regions deemed as suspicious by a computer algorithm and by expert mammographers. Incorporation of the Gaussian saliency map resulted in consistent and often significant improvement of IT-CADe performance across all but one datasets.

Our previously presented information-theoretic computer-aided detection (IT-CADe) system for distinguishing masses and normal parenchyma in mammograms is an example of a case-based system. IT-CAD makes decisions by evaluating the querys average similarity with known mass and normal examples stored in the systems case base. Pairwise case similarity is measured in terms of their normalized mutual information. The purpose of this study was to evaluate whether incorporating a new machine learning concept of relational representation to IT-CAD is a more effective strategy than the decision algorithm that is currently in place. A trainable relational representation classifier builds a decision rule using the relational representation of cases. Instead of describing a case by a vector of intrinsic features, the case is described by its NMI-based similarity to a set of known examples. For this study, we first applied random mutation hill climbing algorithm to select the concise set of knowledge cases and then we applied a support vector machine to derive a decision rule using the relational representation of cases. We performed the study with a database of 600 mammographic regions of interest (300 with masses and 300 with normal parenchyma). Our experiments indicate that incorporating the concept of relational representation with a trainable classifier to IT-CAD provides an improvement in performance as compared with the original decision rule. Therefore, relational representation is a promising strategy for IT-CADe.

Electronic colon cleansing (ECC) aims to remove the contrast agent from the CT abdominal images so that a virtual model of the colon can be constructed. Virtual colonoscopy requires either liquid or solid preparation of the colon before CT imaging. This paper has two parts to address ECC in both preparation methods. In the first part, meniscus removal in the liquid preparation is studied. The meniscus is the curve seen at the top of a liquid in response to its container. Left on the colon wall, the meniscus can decrease the sensitivity and specificity of virtual colonoscopy. We state the differential equation that governs the profile of the meniscus and propose an algorithm for calculating the boundary of the contrast agent. We compute the surface tension of the liquid-colon wall contact using in-vivo CT data. Our results show that the surface tension can be estimated with an acceptable degree of uncertainty. Such an estimate, along with the meniscus profile differential equation will be used as an a priori knowledge to aid meniscus segmentation. In the second part, we study ECC in solid preparation of colon. Since the colon is pressurized with air before acquisition of the CT images, a prior on the shape of the colon wall can be obtained. We present such prior and investigate it using patient data. We show the shape prior is held in certain parts of the colon and propose a method that uses this prior to ease pseudoenhancement correction.

Our purpose was to develop a computer-aided diagnostic (CAD) scheme for detection of flat lesions (also known as superficial elevated or depressed lesions) in CT colonography (CTC), which utilized 3D massive-training artificial neural networks (MTANNs) for false-positive (FP) reduction. Our CAD scheme consisted of colon segmentation, polyp candidate detection, linear discriminant analysis, and MTANNs. To detect flat lesions, we developed a precise shape analysis in the polyp detection step to accommodate the analysis to include a flat shape. With our MTANN CAD scheme, 68% (19/28) of flat lesions, including six lesions "missed" by radiologists in a multicenter clinical trial, were detected correctly, with 10 (249/25) FPs per patient.

CT colonography (CTC) is a feasible and minimally invasive method for the detection of colorectal polyps and cancer screening. Computer-aided detection (CAD) of polyps has improved consistency and sensitivity of virtual colonoscopy interpretation and reduced interpretation burden. A CAD system typically consists of four stages: (1) image preprocessing including colon segmentation; (2) initial detection generation; (3) feature selection; and (4) detection classification. In our experience, three existing problems limit the performance of our current CAD system. First, highdensity orally administered contrast agents in fecal-tagging CTC have scatter effects on neighboring tissues. The scattering manifests itself as an artificial elevation in the observed CT attenuation values of the neighboring tissues. This pseudo-enhancement phenomenon presents a problem for the application of computer-aided polyp detection, especially when polyps are submerged in the contrast agents. Second, general kernel approach for surface curvature computation in the second stage of our CAD system could yield erroneous results for thin structures such as small (6-9 mm) polyps and for touching structures such as polyps that lie on haustral folds. Those erroneous curvatures will reduce the sensitivity of polyp detection. The third problem is that more than 150 features are selected from each polyp candidate in the third stage of our CAD system. These high dimensional features make it difficult to learn a good decision boundary for detection classification and reduce the accuracy of predictions. Therefore, an improved CAD system for polyp detection in CTC data is proposed by introducing three new techniques. First, a scale-based scatter correction algorithm is applied to reduce pseudo-enhancement effects in the image pre-processing stage. Second, a cubic spline interpolation method is utilized to accurately estimate curvatures for initial detection generation. Third, a new dimensionality reduction classifier, diffusion map and local linear embedding (DMLLE), is developed for classification and false positives (FP) reduction. Performance of the improved CAD system is evaluated and compared with our existing CAD system (without applying those techniques) using CT scans of 1186 patients. These scans are divided into a training set and a test set. The sensitivity of the improved CAD system increased 18% on training data at a rate of 5 FPs per patient and 15% on test data at a rate of 5 FPs per patient. Our results indicated that the improved CAD system achieved significantly better performance on medium-sized colonic adenomas with higher sensitivity and lower FP rate in CTC.

One of primary challenges in the medical image data analysis is the ability to handle abnormal, irregular and/or partial cases. In this paper, we present two different robust algorithms towards the goal of automatic planar primitive detection in 3D volumes. The overall algorithm is a bottoms-up approach starting with the anatomic point primitives (or landmarks) detection. The robustness in computing the planar primitives is built in through both a novel consensus-based voting approach, and a random sampling-based weighted least squares regression method. Both these approaches remove inconsistent landmarks and outliers detected in the landmark detection step. Unlike earlier approaches focused towards a particular plane, the presented approach is generic and can be easily adapted to computing more complex primitives such as ROIs or surfaces. To demonstrate the robustness and accuracy of our approach, we present extensive results for automatic plane detection (Mig-Sagittal and Optical Triangle planes) in brain MR-images. In comparison to ground truth, our approach has marginal errors on about 90 patients. The algorithm also works really well under adverse conditions of arbitrary rotation and cropping of the 3D volume. In order to exhibit generalization of the approach, we also present preliminary results on intervertebrae-plane detection for 3D spine MR application.

To make accurate decisions based on imaging data, radiologists must associate the viewed imaging data with the corresponding anatomical structures. Furthermore, given a disease hypothesis possible image findings which verify the hypothesis must be considered and where and how they are expressed in the viewed images. If rare anatomical variants, rare pathologies, unfamiliar protocols, or ambiguous findings are present, external knowledge sources such as medical encyclopedias are consulted. These sources are accessed using keywords typically describing anatomical structures, image findings, pathologies. In this paper we present our vision of how a patient's imaging data can be automatically enhanced with anatomical knowledge as well as knowledge about image findings. On one hand, we propose the automatic annotation of the images with labels from a standard anatomical ontology. These labels are used as keywords for a medical encyclopedia such as STATdx to access anatomical descriptions, information about pathologies and image findings. On the other hand we envision encyclopedias to contain links to region- and finding-specific image processing algorithms. Then a finding is evaluated on an image by applying the respective algorithm in the associated anatomical region. Towards realization of our vision, we present our method and results of automatic annotation of anatomical structures in 3D MRI brain images. Thereby we develop a complex surface mesh model incorporating major structures of the brain and a model-based segmentation method. We demonstrate the validity by analyzing the results of several training and segmentation experiments with clinical data focusing particularly on the visual pathway.

To evaluate the accuracy of computer aided differential diagnosis (CADD) between usual interstitial pneumonia (UIP) and nonspecific interstitial pneumonia (NSIP) at HRCT in comparison with that of a radiologist's decision. A computerized classification for six local disease patterns (normal, NL; ground-glass opacity, GGO; reticular opacity, RO; honeycombing, HC; emphysema, EM; and consolidation, CON) using texture/shape analyses and a SVM classifier at HRCT was used for pixel-by-pixel labeling on the whole lung area. The mode filter was applied on the results to reduce noise. Area fraction (AF) of each pattern, directional probabilistic density function (pdf) (dPDF: mean, SD, skewness of pdf /3 directions: superior-inferior, anterior-posterior, central-peripheral), regional cluster distribution pattern (RCDP: number, mean, SD of clusters, mean, SD of centroid of clusters) were automatically evaluated. Spatially normalized left and right lungs were evaluated separately. Disease division index (DDI) on every combination of AFs and asymmetric index (AI) between left and right lung ((left-right)/left) were also evaluated. To assess the accuracy of the system, fifty-four HRCT data sets in patients with pathologically diagnosed UIP (n=26) and NSIP (n=28) were used. For a classification procedure, a CADD-SVM classifier with internal parameter optimization, and sequential forward floating feature selection (SFFS) were employed. The accuracy was assessed by a 5-folding cross validation with 20- times repetition. For comparison, two thoracic radiologists reviewed the whole HRCT images without clinical information and diagnose each case either as UIP or NSIP. The accuracies of radiologists' decision were 0.75 and 0.87, respectively. The accuracies of the CADD system using the features of AF, dPDF, AI of dPDF, RDP, AI of RDP, DDI were 0.70, 0.79, 0.77, 0.80, 0.78, 0.81, respectively. The accuracy of optimized CADD using all features after SFFS was 0.91. We developed the CADD system to differentiate between UIP and NSIP using automated assessment of the extent and distribution of regional disease patterns at HRCT.

Separating the retinal vascular tree into arteries and veins is important for quantifying vessel changes that preferentially affect either the veins or the arteries. For example the ratio of arterial to venous diameter, the retinal a/v ratio, is well established to be predictive of stroke and other cardiovascular events in adults, as well as the staging of retinopathy of prematurity in premature infants. This work presents a supervised, automatic method that can determine whether a vessel is an artery or a vein based on intensity and derivative information. After thinning of the vessel segmentation, vessel crossing and bifurcation points are removed leaving a set of vessel segments containing centerline pixels. A set of features is extracted from each centerline pixel and using these each is assigned a soft label indicating the likelihood that it is part of a vein. As all centerline pixels in a connected segment should be the same type we average the soft labels and assign this average label to each centerline pixel in the segment. We train and test the algorithm using the data (40 color fundus photographs) from the DRIVE database1 with an enhanced reference standard. In the enhanced reference standard a fellowship trained retinal specialist (MDA) labeled all vessels for which it was possible to visually determine whether it was a vein or an artery. After applying the proposed method to the 20 images of the DRIVE test set we obtained an area under the receiver operator characteristic (ROC) curve of 0.88 for correctly assigning centerline pixels to either the vein or artery classes.

Along with the advancing of technology in wireless and miniature camera, Wireless Capsule Endoscopy (WCE), the combination of both, enables a physician to diagnose patient's digestive system without actually perform a surgical procedure. Although WCE is a technical breakthrough that allows physicians to visualize the entire small bowel noninvasively, the video viewing time takes 1 - 2 hours. This is very time consuming for the gastroenterologist. Not only it sets a limit on the wide application of this technology but also it incurs considerable amount of cost. Therefore, it is important to automate such process so that the medical clinicians only focus on interested events. As an extension from our previous work that characterizes the motility of digestive tract in WCE videos, we propose a new assessment system for energy based events detection (EG-EBD) to classify the events in WCE videos. For the system, we first extract general features of a WCE video that can characterize the intestinal contractions in digestive organs. Then, the event boundaries are identified by using High Frequency Content (HFC) function. The segments are classified into WCE event by special features. In this system, we focus on entering duodenum, entering cecum, and active bleeding. This assessment system can be easily extended to discover more WCE events, such as detailed organ segmentation and more diseases, by using new special features. In addition, the system provides a score for every WCE image for each event. Using the event scores, the system helps a specialist to speedup the diagnosis process.

This paper explores a novel approach to interactive user-guided image segmentation, using eyegaze information as an input. The method includes three steps: 1) eyegaze tracking for providing user input, such as setting object and background seed pixel selection; 2) an optimization method for image labeling that is constrained or affected by user input; and 3) linking the two previous steps via a graphical user interface for displaying the images and other controls to the user and for providing real-time visual feedback of eyegaze and seed locations, thus enabling the interactive segmentation procedure. We developed a new graphical user interface supported by an eyegaze tracking monitor to capture the user's eyegaze movement and fixations (as opposed to traditional mouse moving and clicking). The user simply looks at different parts of the screen to select which image to segment, to perform foreground and background seed placement and to set optional segmentation parameters. There is an eyegaze-controlled "zoom" feature for difficult images containing objects with narrow parts, holes or weak boundaries. The image is then segmented using the random walker image segmentation method. We performed a pilot study with 7 subjects who segmented synthetic, natural and real medical images. Our results show that getting used the new interface takes about only 5 minutes. Compared with traditional mouse-based control, the new eyegaze approach provided a 18.6% speed improvement for more than 90% of images with high object-background contrast. However, for low contrast and more difficult images it took longer to place seeds using the eyegaze-based "zoom" to relax the required eyegaze accuracy of seed placement.

Computer-aided Diagnosis (CAD) systems for the automatic identification of abnormalities in retinal images are gaining importance in diabetic retinopathy screening programs. A huge amount of retinal images are collected during these programs and they provide a starting point for the design of machine learning algorithms. However, manual annotations of retinal images are scarce and expensive to obtain. This paper proposes a dynamic CAD system based on active learning for the automatic identification of hard exudates, cotton wool spots and drusen in retinal images. An uncertainty sampling method is applied to select samples that need to be labeled by an expert from an unlabeled set of 4000 retinal images. It reduces the number of training samples needed to obtain an optimum accuracy by dynamically selecting the most informative samples. Results show that the proposed method increases the classification accuracy compared to alternative techniques, achieving an area under the ROC curve of 0.87, 0.82 and 0.78 for the detection of hard exudates, cotton wool spots and drusen, respectively.

The accurate localization of the optic cup in retinal images is important to assess the cup to disc ratio (CDR) for glaucoma screening and management. Glaucoma is physiologically assessed by the increased excavation of the optic cup within the optic nerve head, also known as the optic disc. The CDR is thus an important indicator of risk and severity of glaucoma. In this paper, we propose a method of determining the cup boundary using non-stereographic retinal images by the automatic detection of a morphological feature within the optic disc known as kinks. Kinks are defined as the bendings of small vessels as they traverse from the disc to the cup, providing physiological validation for the cup boundary. To detect kinks, localized patches are first generated from a preliminary cup boundary obtained via level set. Features obtained using edge detection and wavelet transform are combined using a statistical approach rule to identify likely vessel edges. The kinks are then obtained automatically by analyzing the detected vessel edges for angular changes, and these kinks are subsequently used to obtain the cup boundary. A set of retinal images from the Singapore Eye Research Institute was obtained to assess the performance of the method, with each image being clinically graded for the CDR. From experiments, when kinks were used, the error on the CDR was reduced to less than 0.1 CDR units relative to the clinical CDR, which is within the intra-observer variability of 0.2 CDR units.

This paper presents an approach to the computer aided diagnosis (CAD) of diabetic retinopathy (DR) -- a common and severe complication of long-term diabetes which damages the retina and cause blindness. Since red lesions are regarded as the first signs of DR, there has been extensive research on effective detection and localization of these abnormalities in retinal images. In contrast to existing algorithms, a new approach based on Multiscale Correlation Filtering (MSCF) and dynamic thresholding is developed. This consists of two levels, Red Lesion Candidate Detection (coarse level) and True Red Lesion Detection (fine level). The approach was evaluated using data from Retinopathy On-line Challenge (ROC) competition website and we conclude our method to be effective and efficient.

Diabetic Retinopathy is one of the leading causes of blindness and vision defects in developed countries. An early detection and diagnosis is crucial to avoid visual complication. Microaneurysms are the first ocular signs of the presence of this ocular disease. Their detection is of paramount importance for the development of a computer-aided diagnosis technique which permits a prompt diagnosis of the disease. However, the detection of microaneurysms in retinal images is a difficult task due to the wide variability that these images usually present in screening programs. We propose a statistical approach based on mixture model-based clustering and logistic regression which is robust to the changes in the appearance of retinal fundus images. The method is evaluated on the public database proposed by the Retinal Online Challenge in order to obtain an objective performance measure and to allow a comparative study with other proposed algorithms.

The presence of microaneurysms in the eye is one of the early signs of diabetic retinopathy, which is one of the leading causes of vision loss. We have been investigating a computerized method for the detection of microaneurysms on retinal fundus images, which were obtained from the Retinopathy Online Challenge (ROC) database. The ROC provides 50 training cases, in which "gold standard" locations of microaneurysms are provided, and 50 test cases without the gold standard locations. In this study, the computerized scheme was developed by using the training cases. Although the results for the test cases are also included, this paper mainly discusses the results for the training cases because the "gold standard" for the test cases is not known. After image preprocessing, candidate regions for microaneurysms were detected using a double-ring filter. Any potential false positives located in the regions corresponding to blood vessels were removed by automatic extraction of blood vessels from the images. Twelve image features were determined, and the candidate lesions were classified into microaneurysms or false positives using the rule-based method and an artificial neural network. The true positive fraction of the proposed method was 0.45 at 27 false positives per image. Forty-two percent of microaneurysms in the 50 training cases were considered invisible by the consensus of two co-investigators. When the method was evaluated for visible microaneurysms, the sensitivity for detecting microaneurysms was 65% at 27 false positives per image. Our computerized detection scheme could be improved for helping ophthalmologists in the early diagnosis of diabetic retinopathy.

In recent years, high resolution CT has been developed. CAD system is indispensable for pulmonary cancer screening. In research and development of computer-aided differential diagnosis, there is now widespread interest in the use of nodule doubling time for measuring the volumetric changes of pulmonary nodule. The evolution pattern of each nodule might depend on the CT number distribution pattern inside nodule such as pure GGO, mixed GGO, or solid nodules. This paper presents a computerized approach to measure nodule CT number variation inside pulmonary nodule. The approach consists of four steps: (1) nodule segmentation, (2) computation of CT number histogram, (3) nodule categorization (α, β, γ, ε) based on CT number histogram, (4) computation of doubling time based on CT number histogram, and growth-pattern classification which consists of six categories such as decrease, gradual decrease, no change, slow increase, gradual increase, and increase, and (5) classification between benign and malignant cases. Using our dataset of follow-up scans for whom the final diagnosis was known (62 benign and 42 malignant cases), we evaluated growth-pattern of nodules and designed the classification strategy between benign and malignant cases. In order to compare the performance between the proposed features and volumetric doubling time, the classification result was analyzed by an area under the receiver operating characteristic curve. The preliminary experimental result demonstrated that our approach has a highly potential usefulness to assess the nodule evolution using 3-D thoracic CT images.

CT scanners often have higher in-plane resolution than axial resolution. As a result, measurements in the axial direction are less reliable than measurements in-plane, and this should be considered when performing nodule growth measurements. We propose a method to measure nodule growth rates by a moment-based algorithm using the central second order moments for the in-plane directions. The interscan repeatability of the new method was compared to a volumetric measurement method on a database of 22 nodules with multiple scans taken in the same session. The interscan variability was defined as the 95% confidence interval of the relative volume change. For the entire database of nodules, the interscan variability of the volumetric growth method was (-52.1%, 30.1%); the moment-based method improved the variability to (-34.2%, 23.3%). For the 11 nodules with scans of the same slice thickness between scans, the variability of the volumetric growth method was (24.0%, 30.1%), compared to (-12.4%, 12.7%) for the moment-based method. The 11 nodules with scans of different slice thickness had a variability for the volumetric method of (-68.4%, 30.2%) and for the moment-based method, (-46.5%, 24.4%). The moment-based method showed improvement in interscan variability for all cases. This study shows promising preliminary results of improved repeatability of the new moment-based method over a volumetric method and suggests that measurements on scans of the same slice thickness are more repeatable than on scans of different slice thickness. The 11 nodules with the same slice thickness are publicly available.

The primary stage of a pulmonary nodule detection system is typically a candidate generator that efficiently provides the centroid location and size estimate of candidate nodules. A scale-normalized Laplacian of Gaussian (LOG) filtering method presented in this paper has been found to provide high sensitivity along with precise locality and size estimation. This approach involves a computationally efficient algorithm that is designed to identify all solid nodules in a whole lung anisotropic CT scan. This nodule candidate generator has been evaluated in conjunction with a set of discriminative features that target both isolated and attached nodules. The entire detection system was evaluated with respect to a sizeenriched dataset of 656 whole-lung low-dose CT scans containing 459 solid nodules with diameter greater than 4 mm. Using a soft margin SVM classifier, and setting false positive rate of 10 per scan, we obtained a sensitivity of 97% for isolated, 93% for attached, and 89% for both nodule types combined. Furthermore, the LOG filter was shown to have good agreement with the radiologist ground truth for size estimation.

The computer-aided detection (CAD) system we applied on the ANODE09 dataset is devoted to identify pulmonary nodules in low-dose and thin-slice computed tomography (CT) images: we developed two different systems for internal (CAD_I) and juxtapleural nodules (CAD_JP) in the framework of the italian MAGIC-5 collaboration. The basic modules of CADI subsystem are: a 3D dot-enhancement algorithm for nodule candidate identification and an original approach, we referred as Voxel-Based Neural Approach (VBNA), to reduce the amount of false-positive findings based on a neural classifier working at the voxel level. To detect juxtapleural nodules we developed the CADJP subsystem based on a procedure enhancing regions where many pleura surface normals intersect, followed by a VBNA classification. We present both the FROC curves we obtained on the 5 annotated ANODE09 example dataset, and on all the ANODE09 50 test cases.

A computer-aided detection (CAD) system for lung nodules in CT scans was developed. For the detection of lung nodules two different methods were applied and only pixels which were detected by both methods are marked as true positives. The first method uses a multi-threshold algorithm, which detect connected regions within the lung that have an intensity between specified threshold values. The second is a multi-scale detection method. The data are searched for points located in spherical objects. The image data were smoothed with a 3D Gaussian filter and computed the Hessian matrix and eigenvectors and eigenvalues for all pixels detected by the first algorithm. By analyzing the eigenvalues points that lie within a spherical structure can be located. For segmentation of the detected nodules an active contour model was used. A two-dimensional active contour with four energy terms describing form and position of the contour in the image data was implemented. In addition balloon energy to get the active contour was used growing out from one point. The result of our detection part is used as input for the segmentation part. To test the detection algorithms we used 19 CT volume data sets from a low-dose CT studies. Our CAD system detected 58% of the nodules with a falsepositive rate of 1.38. Additionally we take part at the ANODE09 study whose results will be presented at the SPIE meeting in 2009.

The Lung Test Images from Motol Environment (Lung TIME) is a new publicly available dataset of thoracic CT scans with manually annotated pulmonary nodules. It is larger than other publicly available datasets. Pulmonary nodules are lesions in the lungs, which may indicate lung cancer. Their early detection significantly improves survival rate of patients. Automatic nodule detecting systems using CT scans are being developed to reduce physicians' load and to improve detection quality. Besides presenting our own nodule detection system, in this article, we mainly address the problem of testing and comparison of automatic nodule detection methods. Our publicly available 157 CT scan dataset with 394 annotated nodules contains almost every nodule types (pleura attached, vessel attached, solitary, regular, irregular) with 2-10mm in diameter, except ground glass opacities (GGO). Annotation was done consensually by two experienced radiologists. The data are in DICOM format, annotations are provided in XML format compatible with the Lung Imaging Database Consortium (LIDC). Our computer aided diagnosis system (CAD) is based on mathematical morphology and filtration with a subsequent classification step. We use Asymmetric AdaBoost classifier. The system was tested using TIME, LIDC and ANODE09 databases. The performance was evaluated by cross-validation for Lung TIME and LIDC, and using the supplied evaluation procedure for ANODE09. The sensitivity at chosen working point was 94.27% with 7.57 false positives/slice for TIME and LIDC datasets combined, 94.03% with 5.46 FPs/slice for the Lung TIME, 89.62% sensitivity with 12.03 FPs/slice for LIDC, and 78.68% with 4,61 FPs/slice when applied on ANODE09.

Labeled training data in the medical domain is rare and expensive to obtain. The lack of labeled multimodal medical image data is a major obstacle for devising learning-based interactive segmentation tools. Transductive learning (TL) or semi-supervised learning (SSL) offers a workaround by leveraging unlabeled and labeled data to infer labels for the test set given a small portion of label information. In this paper we propose a novel algorithm for interactive segmentation using transductive learning and inference in conditional mixture naïve Bayes models (T-CMNB) with spatial regularization constraints. T-CMNB is an extension of the transductive naïve Bayes algorithm [1, 20]. The multimodal Gaussian mixture assumption on the class-conditional likelihood and spatial regularization constraints allow us to explain more complex distributions required for spatial classification in multimodal imagery. To simplify the estimation we reduce the parameter space by assuming naïve conditional independence between the feature space and the class label. The naïve conditional independence assumption allows efficient inference of marginal and conditional distributions for large scale learning and inference [19]. We evaluate the proposed algorithm on multimodal MRI brain imagery using ROC statistics and provide preliminary results. The algorithm shows promising segmentation performance with a sensitivity and specificity of 90.37% and 99.74% respectively and compares competitively to alternative interactive segmentation schemes.

Diffusion weighted imaging (DWI) derived apparent diffusion coefficient (ADC) values are known to correlate inversely to tumour cellularity in brain tumours. The average ADC value increases after successful chemotherapy, radiotherapy or a combination of both and can be therewith used as a surrogate marker for treatment response. Moreover, high and low malignant areas can be distinguished. The main purpose of our project was to develop a software platform that enables the automated delineation and ADC quantification of different tumour sections in a fast, objective, user independent manner. Moreover, the software platform allows for an analysis of the probability density of the ADC in high and low malignant areas in ROIs drawn on conventional imaging to create a ground truth. We tested an Expectation Maximization algorithm with a Gaussian mixture model to objectively determine tumour heterogeneity in gliomas because of yielding Gaussian distributions in the different areas. Furthermore, the algorithm was initialized by seed points in the areas of the gross tumour volume and the data indicated that an automatic initialization should be possible. Thus automated clustering of high and low malignant areas and subsequent ADC determination within these areas is possible yielding reproducible ADC measurements within heterogeneous gliomas.

This study applied a Gaussian Mixture Model (GMM) to apparent diffusion coefficient (ADC) histograms to evaluate glioblastoma multiforme (GBM) tumor treatment response using diffusion weighted (DW) MR images. ADC mapping, calculated from DW images, has been shown to reveal changes in the tumor's microenvironment preceding morphologic tumor changes. In this study, we investigated the effectiveness of features that represent changes from pre- and post-treatment tumor ADC histograms to detect treatment response. The main contribution of this work is to model the ADC histogram as the composition of two components, fitted by GMM with expectation maximization (EM) algorithm. For both pre- and post-treatment scans taken 5-7 weeks apart, we obtained the tumor ADC histogram, calculated the two-component features, as well as the other standard histogram-based features, and applied supervised learning for classification. We evaluated our approach with data from 85 patients with GBM under chemotherapy, in which 33 responded and 52 did not respond based on tumor size reduction. We compared AdaBoost and random forests classification algorithms, using ten-fold cross validation, resulting in a best accuracy of 69.41%.

Accurate ventricle volume estimates could potentially improve the understanding and diagnosis of communicating hydrocephalus. Postoperative communicating hydrocephalus has been recognized in patients with brain tumors where the changes in ventricle volume can be difficult to identify, particularly over short time intervals. Because of the complex alterations of brain morphology in these patients, the segmentation of brain ventricles is challenging. Our method evaluates ventricle size from serial brain MRI examinations; we (i) combined serial images to increase SNR, (ii) automatically segmented this image to generate a ventricle template using fast marching methods and geodesic active contours, and (iii) propagated the segmentation using deformable registration of the original MRI datasets. By applying this deformation to the ventricle template, serial volume estimates were obtained in a robust manner from routine clinical images (0.93 overlap) and their variation analyzed.

Our previous works suggest that fractal-based texture features are very useful for detection, segmentation and classification of posterior-fossa (PF) pediatric brain tumor in multimodality MRI. In this work, we investigate and compare efficacy of our texture features such as fractal and multifractional Brownian motion (mBm), and intensity along with another useful level-set based shape feature in PF tumor segmentation. We study feature selection and ranking using Kullback -Leibler Divergence (KLD) and subsequent tumor segmentation; all in an integrated Expectation Maximization (EM) framework. We study the efficacy of all four features in both multimodality as well as disparate MRI modalities such as T1, T2 and FLAIR. Both KLD feature plots and information theoretic entropy measure suggest that mBm feature offers the maximum separation between tumor and non-tumor tissues in T1 and FLAIR MRI modalities. The same metrics show that intensity feature offers the maximum separation between tumor and non-tumor tissue in T2 MRI modality. The efficacies of these features are further validated in segmenting PF tumor using both single modality and multimodality MRI for six pediatric patients with over 520 real MR images.

Computer systems play an important role in medical imaging industry since radiologists depend on it for visualization, interpretation, communication and archiving. In particular, computer-aided diagnosis (CAD) systems help in lesion detection tasks. This paper presents the design and the development of an interactive segmentation tool for breast cancer screening and diagnosis. The tool conception is based upon a user-centered approach in order to ensure that the application is of real benefit to radiologists. The analysis of user expectations, workflow and decision-making practices give rise to the need for an interactive reporting system based on the BIRADS, that would not only include the numerical features extracted from the segmentation of the findings in a structured manner, but also support human relevance feedback as well. This way, the numerical results from segmentation can be either validated by end-users or enhanced thanks to domain-experts subjective interpretation. Such a domain-expert centered system requires the segmentation to be sufficiently accurate and locally adapted, and the features to be carefully selected in order to best suit user's knowledge and to be of use in enhancing segmentation. Improving segmentation accuracy with relevance feedback and providing radiologists with a user-friendly interface to support image analysis are the contributions of this work. The preliminary result is first the tool conception, and second the improvement of the segmentation precision.

Mammographic breast density is a known breast cancer risk factor. Studies have shown the potential to automate breast density estimation by using computerized texture-based segmentation of the dense tissue in mammograms. Digital breast tomosynthesis (DBT) is a tomographic x-ray breast imaging modality that could allow volumetric breast density estimation. We evaluated the feasibility of distinguishing between dense and fatty breast regions in DBT using computer-extracted texture features. Our long-term hypothesis is that DBT texture analysis can be used to develop 3D dense tissue segmentation algorithms for estimating volumetric breast density. DBT images from 40 women were analyzed. The dense tissue area was delineated within each central source projection (CSP) image using a thresholding technique (Cumulus, Univ. Toronto). Two (2.5cm)² ROIs were manually selected: one within the dense tissue region and another within the fatty region. Corresponding (2.5cm)³ ROIs were placed within the reconstructed DBT images. Texture features, previously used for mammographic dense tissue segmentation, were computed. Receiver operating characteristic (ROC) curve analysis was performed to evaluate feature classification performance. Different texture features appeared to perform best in the 3D reconstructed DBT compared to the 2D CSP images. Fractal dimension was superior in DBT (AUC=0.90), while contrast was best in CSP images (AUC=0.92). We attribute these differences to the effects of tissue superimposition in CSP and the volumetric visualization of the breast tissue in DBT. Our results suggest that novel approaches, different than those conventionally used in projection mammography, need to be investigated in order to develop DBT dense tissue segmentation algorithms for estimating volumetric breast density.

Computer aided diagnosis (CADx) systems have the potential to support the radiologist in the complex task of discriminating benign and malignant types of breast lesions based on their appearance in mammograms. Previously, we have proposed a knowledge-based CADx approach for mammographic mass lesions using case-based reasoning. The input of the systems reasoning process are features that are automatically extracted from regions of interest (ROIs) depicting mammographic masses. However, despite the fact that the shape of a mass as well as the characteristics of its boundary are highly discriminative attributes for its diagnosis, we have not included shape and boundary features that are based on an explicit segmentation of the mass from the background tissue in the previously proposed CADx approach. Hence, we present a novel method for the segmentation of mammographic masses in this work and describe how we have integrated this segmentation module into our existent CADx system. The approach is based on the observation that the optical density of a mass is usually high near its core and decreases towards its boundary. Because of tissue superposition and the broad variety of appearances of masses, their automatic segmentation is a difficult task. Thus, it is not surprising that even after many years of research concerning the segmentation of masses no fully automatic approach that robustly solves the problem seems to exist. For this reason, we have included optional interactive modules in the proposed segmentation approach that allow fast and easy corrective interference of the radiologist with the segmentation process.

Dynamic Contrast Enhanced MRI (DCE-MRI) has today a well-established role, complementary to routine imaging techniques for breast cancer diagnosis such as mammography. Despite its undoubted clinical advantages, DCE-MRI data analysis is time-consuming and Computer Aided Diagnosis (CAD) systems are required to help radiologists. Segmentation is one of the key step of every CAD image processing pipeline, but most techniques available require human interaction. We here present the preliminary results of a fully automatic lesion detection method, capable of dealing with fat suppression image acquisition sequences, which represents a challenge for image processing algorithms due to the low SNR. The method is based on four fundamental steps: registration to correct for motion artifacts; anatomical segmentation to discard anatomical structures located outside clinically interesting lesions; lesion detection to select enhanced areas and false positive reduction based on morphological and kinetic criteria. The testing set was composed by 13 cases and included 27 lesions (10 benign and 17 malignant) of diameter > 5 mm. The system achieves a per-lesion sensitivity of 93%, while yielding an acceptable number of false positives (26 on average). The results of our segmentation algorithm were verified by visual inspection, and qualitative comparison with a manual segmentation yielded encouraging results.

Cone Beam Breast CT (CBBCT) acquires 3D breast images without compression to the breast. More detailed and accurate information of breast lesions is revealed in CBBCT images. In our research, based on the observation that tumor masses are more concentrated than the surrounding tissues, we designed a weighted average filter and a threedimensional Iris filter to operate on the three-dimensional images. The basic process is: After weighted average filtering and iris filtering, a thresholding is applied to extract suspicious regions. Next, after morphological processing, suspicious regions are sorted based on their average Iris filter responses and the top 10 candidates are selected as detection results. The detection results are marked out and provided to radiologists as CAD system output. In our experiment, our method detects 12 mass locations out of 14 pathology-proven malignant clinical cases.

The amplification status of human epidermal growth factor receptors 2 (HER2) genes is strongly associated with clinical outcome in patients with breast cancer. The American Society of Clinical Oncology Tumor Marker Guidelines Panel has recommended routine testing of HER2 status on all newly diagnosed metastatic breast cancers since 2001. Although fluorescent in situ hybridization (FISH) technology provides superior accuracy as compared with other approaches, current manual FISH analysis methods are somewhat subjective, tedious, and may introduce interreader variability. The goal of this preliminary study is to develop and test a computer-aided detection (CAD) scheme to assess HER2 status using FISH images. Forty FISH images were selected for this study from our genetic laboratory. The CAD scheme first applies an adaptive, iterative threshold method followed by a labeling algorithm to segment cells of possible interest. A set of classification rules is then used to identify analyzable interphase cells and discard nonanalyzable cells due to cell overlapping and/or other image staining debris (or artifacts). The scheme then maps the detected analyzable cells onto two other gray scale images corresponding to the red and green color of the original image followed by application of a raster scan and labeling algorithms to separately detect the HER-2/neu ("red") and CEP17 ("green") FISH signals. A simple distance based criterion is applied to detect and merge split FISH signals within each cell. The CAD scheme computes the ratio between independent "red" and "green" FISH signals of all analyzable cells identified on an image. If the ratio is ≥ 2.0, the FISH image is assumed to have been acquired from a HER2+ case; otherwise, the FISH image is assumed to have been acquired from HER2- case. When we applied the CAD scheme to the testing dataset, the average computed HER2 amplification ratios were 1.06±0.25 and 2.53±0.81 for HER2- and HER2+ samples, respectively. The results show that the CAD scheme has the ability to automatically detect HER2 status using FISH images. The success of CAD-guided FISH image analysis could result in a more objective, consistent, and efficient approach in determining HER2 status of breast cancers.

Dynamic Contrast Enhanced Breast MR Imaging (DCE BMRI) has emerged as a modality for breast cancer diagnosis. In this modality a temporal sequence of volume images of the breasts is acquired, where a contrast agent is injected after acquisition of the first 3D image. Since the introduction of the modality, research has been directed at the development of computer-aided support for the diagnostic workup. This includes automatic segmentation of mass-like lesions, lesion characterization, and lesion classification. Robustness, user-independence, and reproducibility of the results of computerized methods are essential for such methods to be acceptable for clinical application. A previously proposed and evaluated computerized lesion segmentation method has been further analyzed in this study. The segmentation method uses as input a subtraction image (post-contrast - pre-contrast) and a user defined region of interest (ROI). Previous evaluation studies investigated the robustness of the segmentation against variations in the user selected ROI. Robustness of the method against variations in the image data itself has so far not been investigated. To fill this gap is the purpose of this study. In this study, the segmentation algorithm was applied to a series of subtraction images built from the pre-contrast volume and all available post-contrast image volumes, successively. This provides set of typically 4-5 delineations per lesion, each based on a different phase of the dynamic sequence. Analysis of the apparent lesion volumes derived from these delineations and comparison to manual delineations showed that computerized segmentation is more robust and reproducible than manual segmentation, even if computer segmentations are computed on subtraction images derived from different dynamic phases of the DCE MRI study, while all manual segmentations of a lesion are derived from one and the same dynamic phase of the study. Furthermore, it could be shown that the rate of apparent change of lesion volume over the course of a DCE MRI study is significantly dependent on the lesion type (benign vs. malignant).

We investigated three classifiers for the task of distinguishing between benign and malignant breast lesions. Classification performance was measured in terms of area under the ROC curve (AUC value). We compared linear discriminant analysis (LDA), quadratic discriminant analysis (QDA) and a Bayesian neural net (BNN) with 5 hidden units. For each lesion, 46 image features were extracted and principal component analysis (PCA) of these features was used as classifier input. For each classifier, the optimal number of principal components was determined by performing PCA within each step of a leave-one-case-out protocol for the training dataset (1125 lesions, 14% cancer prevalence) and determining which number of components maximized the AUC value. Subsequently, each classifier was trained on the training dataset and applied 'cold turkey' to an independent test set from a different population (341 lesions, 30% cancer prevalence). The optimal number of principal components for LDA was 24, accounting for 97% of the variance in the image features. For QDA and BNN, these numbers were 5 (70%) and 15 (93%), respectively. The LDA, QDA and BNN obtained AUC values of 0.88, 0.85, and 0.91, respectively, in the leave-one-case-out analysis. In the independent test - with AUCs of 0.88, 0.76, and 0.82 - only LDA achieved performance identical to that for the training set (lower bound of 95% non-inferiority interval -.0067), while the others performed significantly worse (p-values << 0.05). While the more complex BNN classifier outperformed the others in leave-one-case-out of a large dataset, LDA was the robust best-performer in an independent test.

Duct patterns are formed by desmoplastic reactions as most breast carcinomas are. Hence, it has been suggested that the denser a breast is, the higher the likelihood to develop breast cancer. Consequently, breast density has been one of the suggested parameters to estimate the risk to develop breast cancer. Currently, the main technique to evaluate breast densities is through mammograms. However, mammograms have the disadvantage of displaying overlapping structures within the breast. Although there are efficient techniques to obtain breast densities from mammograms, mammography can only provide a rough estimate because of the overlapping breast tissue. In this study, cone beam CT images were utilized to evaluate the breast density of sixteen breast images. First, a breast phantom with known volumes representing fatty, glandular and calcified tissues was designed to calibrate the system. Since cone beam CT provides 3D-isotropic resolution images throughout the field of view, the issue of overlapping structures disappears, allowing greater accuracy in evaluating the volumes of each different part of the phantom. Then, using cone beam CT breast images, the breast density of eight patients was evaluated using a semi-automatic segmentation algorithm that differentiates between fatty, glandular and calcified tissues. The results demonstrated that cone beam CT images provide a better tool to evaluate the breast density of the whole breast more accurately. The results also demonstrated that using this semi-automatic segmentation algorithm improves the efficiency of classifying the breast into the four classifications as recommended by the American College of Radiology.

Our goal was to develop and evaluate a reliable segmentation method to delineate axillary lymph node (ALN) from surrounding tissues on US images as the first step of building a multi-modality CADx system for staging ALN. Ultrasound images of 24 ALN from 18 breast cancer patients were used. An elliptical model algorithm was used to fit ALNs boundaries using the following steps: reduce image noise, extract image edges using the Canny edge detector, select edge pixels and fit an ellipse by minimizing the quadratic error, Find the best fitting ellipse based on RANSAC. The segmentation was qualitatively evaluated by 3 expert readers using 4 aspects: Orientation of long axis (OLA): within +- 45 degrees, or off by +-45 degrees, overlap (OV): the fitted ellipse completely included ALN, partially included ALN, or missed the ALN, size (SZ): too small, good within 20% error margin, or too large, and aspect ratio (AR): correct or wrong. Nightly six % of ALNs were correctly evaluated by all readers in terms of OLA and AR, 90.2% in terms of OV and 86.11 in terms of SZ. Readers agreed that the segmentation was correct in 70% of the cases in all aspects. Due to small sample size and small variation among readers, we don't have power to show the accuracy of them is different.

Structural texture measures are used to address the aspect of breast cancer risk assessment in screening mammograms. The current study investigates whether texture properties characterized by local Fractal Dimension (FD) and Lacunarity contribute to asses breast cancer risk. FD represents the complexity while the Lacunarity characterize the gappiness of a fractal. Our cross-sectional case-control study includes mammograms of 50 patients diagnosed with breast cancer in the subsequent 2-4 years and 50 matched controls. The longitudinal double blind placebo controlled HRT study includes 39 placebo and 36 HRT treated volunteers for two years. ROIs with same dimension (250*150 pixels) were created behind the nipple region on these radiographs. Box counting method was used to calculate the fractal dimension (FD) and the Lacunarity. Paired t-test and Pearson correlation coefficient were calculated. It was found that there were no differences between cancer and control group for FD (P=0.8) and Lacunarity (P=0.8) in crosssectional study whereas earlier published heterogeneity examination of radiographs (BC-HER) breast cancer risk score separated groups (p=0.002). In the longitudinal study, FD decreased significantly (P<0.05) in the HRT treated population while Lacunarity remained insignificant (P=0.2). FD is negatively correlated to Lacunarity (-0.74, P<0.001), BIRADS (-0.34, P<0.001) and Percentage Density (-0.41, P<0.001). FD is invariant to the mammographic texture change from control to cancer population but marginally varying in HRT treated population. This study yields no evidence that lacunarity or FD are suitable surrogate markers of mammographic heterogeneity as they neither pick up breast cancer risk, nor show good sensitivity to HRT.

Database characteristics can affect significantly the performance of a mammography CAD scheme. Hence adequate performance comparison among different CAD schemes is not suitable since a single scheme could present different results depending on the set of chosen cases. Images in database should follow a set of quality criteria, since the imaging process up to digital file. CAD schemes can not be developed without a database used to test their efficacy, but each database with particular characteristics may influence on the processing scheme performance. A possible solution could be using information on the imaging equipment characteristics. This work describes a preprocessing in order to "compensate" the image degradation during the acquisition steps, assuring a better "uniformity" relative to the images quality. Thus, poor quality images would be restored, providing therefore some independence on the images source to CAD schemes and allowing to reach the better possible performance. Tests performed with mammography images sets reported a 14% increase in sensitivity for microcalcifications detection. Although this result was followed by a little increase in false positive rates, simple changes in techniques parameters can provide the same improvement but with a reduction of the false positive detections.

We are investigating the feasibility of predicting the risk of developing breast cancer in future years by analysis of breast parenchymal patterns on mammograms. A data set of CC-view mammograms from prior exams of 96 cancer patients and 491 normal subjects was collected from patient files. The prior mammograms were obtained at least one year before diagnosis for cancer patients and two-years of cancer-free follow-up for normal subjects. The percent dense area was estimated by automated gray-level histogram analysis. Texture features were extracted from a region of interest in the retroareolar area. A feature space was constructed by using the percent dense area and texture features in combination with patient age. A linear discriminant analysis (LDA) classifier with stepwise feature selection was trained to evaluate whether the breast parenchyma of future cancer patients can be distinguished from those of normal subjects in the selected feature space. The areas under receiver operating characteristic curves (Az) were 0.90±0.02, 0.86±0.02, and 0.69±0.02 for the classification of future cancer patients and normal subjects by using the breast that would develop cancer (M-vs-N), the contralateral breast (CoM-vs-N) and the patient age only (C-vs-N-with-Age), respectively. The difference in the Az values between the M-vs-N and CoM-vs-N approaches did not achieve statistical significance (p=0.11) by using ROC analysis. The performances of M-vs-N and CoM-vs-N were significantly better than that of C-vs-N-with-Age (p<0.05). Our preliminary result indicates that computerized mammographic parenchymal analysis might be useful for predicting the elevated risk of developing breast cancer in future years.

The aim of our work is to present a robust 3D automated method for measuring regional myocardial thickening using cardiac magnetic resonance imaging (MRI) based on Laplace's equation. Multiple slices of the myocardium in short-axis orientation at end-diastolic and end-systolic phases were considered for this analysis. Automatically assigned 3D epicardial and endocardial boundaries were fitted to short-axis and long axis slices corrected for breathold related misregistration, and final boundaries were edited by a cardiologist if required. Myocardial thickness was quantified at the two cardiac phases by computing the distances between the myocardial boundaries over the entire volume using Laplace's equation. The distance between the surfaces was found by computing normalized gradients that form a vector field. The vector fields represent tangent vectors along field lines connecting both boundaries. 3D thickening measurements were transformed into polar map representation and 17-segment model (American Heart Association) regional thickening values were derived. The thickening results were then compared with standard 17-segment 6-point visual scoring of wall motion/wall thickening (0=normal; 5=greatest abnormality) performed by a consensus of two experienced imaging cardiologists. Preliminary results on eight subjects indicated a strong negative correlation (r=-0.8, p<0.0001) between the average thickening obtained using Laplace and the summed segmental visual scores. Additionally, quantitative ejection fraction measurements also correlated well with average thickening scores (r=0.72, p<0.0001). For segmental analysis, we obtained an overall correlation of -0.55 (p<0.0001) with higher agreement along the mid and apical regions (r=-0.6). In conclusion 3D Laplace transform can be used to quantify myocardial thickening in 3D.

The detection of cerebrovascular diseases such as unruptured aneurysm, stenosis, and occlusion is a major application of magnetic resonance angiography (MRA). However, their accurate detection is often difficult for radiologists. Therefore, several computer-aided diagnosis (CAD) schemes have been developed in order to assist radiologists with image interpretation. The purpose of this study was to develop a computerized method for segmenting cerebral arteries, which is an essential component of CAD schemes. For the segmentation of vessel regions, we first used a gray level transformation to calibrate voxel values. To adjust for variations in the positioning of patients, registration was subsequently employed to maximize the overlapping of the vessel regions in the target image and reference image. The vessel regions were then segmented from the background using gray-level thresholding and region growing techniques. Finally, rule-based schemes with features such as size, shape, and anatomical location were employed to distinguish between vessel regions and false positives. Our method was applied to 854 clinical cases obtained from two different hospitals. The segmentation of cerebral arteries in 97.1%(829/854) of the MRA studies was attained as an acceptable result. Therefore, our computerized method would be useful in CAD schemes for the detection of cerebrovascular diseases in MRA images.

The successful application of Computer Aided Detection schemes to CT Colonography depends not only on their performances in terms of sensitivity and specificity, but also on the interaction with the radiologist, and thus ultimately on factors such as the nature of CAD prompts and the reading paradigm. Fecal tagging is emerging as a widely accepted technique for patient preparation, and patient-friendlier schemes are being proposed in an effort to increase compliance to screening programs; the interaction between CAD and FT regimens should likewise be taken into account. In this scenario, an analysis of the characteristics of CAD prompts is of paramount importance in order to guide further research, both from clinical and technical viewpoints. The CAD scheme analyzed in this paper is essentially composed of five steps: electronic cleansing, colon surface extraction, polyp candidate segmentation, pre-filtering of residual tagged stool and classification of the generated candidates in true polyps vs. false alarms. False positives were divided into six categories: untagged and tagged solid stool, haustral folds, extra-colonic candidates, ileocecal valve and cleansing artifacts. A full cathartic preparation was compared with a semi-cathartic regimen with sameday fecal tagging, which is characterized by higher patient acceptance but also higher inhomogeneity. The distribution of false positives at segmentation reflects the quality of preparation, as more inhomogeneous tagging results in a higher number of untagged solid stool and cleansing artifacts.

This paper presents a haustral fold detection method from 3D abdominal CT images. CT colonography (CTC) or virtual colonoscopy is a new colon diagnostic method to examine the inside of the colon. CTC system can visualize the interior of the colon from any viewpoint and viewing direction based on CT images of a patient. Both the supine and the prone positions of CT images are used for colon diagnosis to improve sensitivity of lesion detection. Registration of the supine and the prone positions of a patient is needed to improve efficiency of diagnosis using CT images in the two positions. Positions of haustral folds are utilizable as landmarks to establish correspondence between these two positions. We present a haustral fold detection method for registration of the supine and the prone positions. Haustral folds protrude almost perpendicular to a centerline of the colon. We designed new difference filter of CT values that can detect haustral folds. The difference filter calculates difference values of CT values along the colon centerline. It outputs high values in the haustral folds. False positive elimination is performed using two feature values including output value of the difference filter and volume of connected component. As the results of experiments using 12 cases of CT images, we confirmed that the proposed method can detect haustral folds satisfactorily. From evaluation using haustral folds ⪆ 3 [mm] in height and thickness, sensitivity of our method was 90.8% with 6.1 FPs/case.

With the development of computer-aided detection of polyps (CADpolyp), various features have been extracted to detect the initial polyp candidates (IPCs). In this paper, three approaches were utilized to reduce the number of false positives (FPs): the multiply linear regression (MLR) and two modified machine learning methods, i.e., neural network (NN) and support vector machine (SVM), based on their own characteristics and specific learning purposes. Compared to MLR, the two modified machine learning methods are much more sophisticated and well-adapted to the data provided. To achieve the optimal sensitivity and specificity, raw features were pre-processed by the principle component analysis (PCA) in the hope of removing the second-order statistical correlation prior to any learning actions. The gain by the use of PCA was evidenced by the collected 26 patient studies, which included 32 colonic polyps confirmed by both optical colonoscopy (OC) and virtual colonoscopy (VC). The learning and testing results showed that the two modified machine-learning methods can reduce the number of FPs by 48.9% (or 7.2 FPs per patient) and 45.3% (or 7.7 FPs per patient) respectively, at 100% detection sensitivity in comparison with that of traditional MLR method. Generally, more than necessary number of features were stacked as input vectors to machine learning algorithms, dimensionality reduction for a more compact feature combination, i.e., how to determine the remaining dimensionality via PCA linear transform was considered and discussed in this paper. In addition, we proposed a new PCA-scaled data pre-processing method to help reduce the FPs significantly. Finally, fROC (free-response receiver operating characteristic) curves corresponding to three FP-reduction approaches were acquired, and comparative analysis was conducted.

Fecal tagging preparations are attracting notable interest as a way to increase patients' compliance to virtual colonoscopy. Patient-friendly preparations, however, often result in less homogeneous tagging. Electronic cleansing algorithms should be capable of dealing with such preparations and yield good quality 2D and 3D images; moreover, successful electronic cleansing lays the basis for the application of Computer Aided Detection schemes. In this work, we present a cleansing algorithm based on an adaptive remapping procedure, which is based on a model of how partial volume affects both the air-tissue and the soft-tissue interfaces. Partial volume at the stool-soft tissue interface is characterized in terms of the local characteristics of tagged regions, in order to account for variations in tagging intensity throughout the colon. The two models are then combined in order to obtain a remapping equation relating the observed intensity to the that of the cleansed colon. The electronic cleansed datasets were then processed by a CAD scheme composed of three main steps: colon surface extraction, polyp candidate segmentation through curvature-based features, and linear classifier-based discrimination between true polyps and false alarms. Results obtained were compared with a previous version of the cleansing algorithm, in which a simpler remapping procedure was used. Performances are increased both in terms of the visual quality of the 2D cleansed images and 3D rendered volumes, and of CAD performances on a sameday FT virtual colonoscopy dataset.

Colorectal cancer is the third leading cause of cancer deaths in the United States of America for both women and men. By means of early detection, the five year survival rate can be up to 90%. Polyps can to be grouped into three different classes: hyperplastic, adenomatous, and carcinomatous polyps. Hyperplastic polyps are benign and are not likely to develop into cancer. Adenomas, on the other hand, are known to grow into cancer (adenoma-carcinoma sequence). Carcinomas are fully developed cancers and can be easily distinguished from adenomas and hyperplastic polyps. A recent narrow band imaging (NBI) study by Tischendorf et al. has shown that hyperplastic polyps and adenomas can be discriminated by their blood vessel structure. We designed a computer-aided system for the differentiation between hyperplastic and adenomatous polyps. Our development aim is to provide the medical practitioner with an additional objective interpretation of the available image data as well as a confidence measure for the classification. We propose classification features calculated on the basis of the extracted blood vessel structure. We use the combined length of the detected blood vessels, the average perimeter of the vessels and their average gray level value. We achieve a successful classification rate of more than 90% on 102 polyps from our polyp data base. The classification results based on these features are compared to the results of Local Binary Patterns (LBP). The results indicate that the implemented features are superior to LBP.

Colon cancer is the second leading cause of cancer-related deaths in the United States. Computed tomographic colonography (CTC) combined with a computer aided detection system provides a feasible combination for improving colonic polyps detection and increasing the use of CTC for colon cancer screening. To distinguish true polyps from false positives, various features extracted from polyp candidates have been proposed. Most of these features try to capture the shape information of polyp candidates or neighborhood knowledge about the surrounding structures (fold, colon wall, etc.). In this paper, we propose a new set of shape descriptors for polyp candidates based on statistical curvature information. These features, called histogram of curvature features, are rotation, translation and scale invariant and can be treated as complementing our existing feature set. Then in order to make full use of the traditional features (defined as group A) and the new features (group B) which are highly heterogeneous, we employed a multiple kernel learning method based on semi-definite programming to identify an optimized classification kernel based on the combined set of features. We did leave-one-patient-out test on a CTC dataset which contained scans from 50 patients (with 90 6-9mm polyp detections). Experimental results show that a support vector machine (SVM) based on the combined feature set and the semi-definite optimization kernel achieved higher FROC performance compared to SVMs using the two groups of features separately. At a false positive per patient rate of 7, the sensitivity on 6-9mm polyps using the combined features improved from 0.78 (Group A) and 0.73 (Group B) to 0.82 (p<=0.01).

The evolution of colon cancer starts with colon polyps. There are two different types of colon polyps, namely hyperplasias and adenomas. Hyperplasias are benign polyps which are known not to evolve into cancer and, therefore, do not need to be removed. By contrast, adenomas have a strong tendency to become malignant. Therefore, they have to be removed immediately via polypectomy. For this reason, a method to differentiate reliably adenomas from hyperplasias during a preventive medical endoscopy of the colon (colonoscopy) is highly desirable. A recent study has shown that it is possible to distinguish both types of polyps visually by means of their vascularization. Adenomas exhibit a large amount of blood vessel capillaries on their surface whereas hyperplasias show only few of them. In this paper, we show the feasibility of computer-based classification of colon polyps using vascularization features. The proposed classification algorithm consists of several steps: For the critical part of vessel segmentation, we implemented and compared two segmentation algorithms. After a skeletonization of the detected blood vessel candidates, we used the results as seed points for the Fast Marching algorithm which is used to segment the whole vessel lumen. Subsequently, features are computed from this segmentation which are then used to classify the polyps. In leave-one-out tests on our polyp database (56 polyps), we achieve a correct classification rate of approximately 90%.

In order to eliminate or weaken the interference between different topological structures on the colon wall, adaptive and normalized convolution methods were used to compute the first and second order spatial derivatives of computed tomographic colonography images, which is the beginning of various geometric analyses. However, the performance of such methods greatly depends on the single-layer representation of the colon wall, which is called the starting layer (SL) in the following text. In this paper, we introduce a level set-based adaptive convolution (LSAC) method to compute the spatial derivatives, in which the level set method is employed to determine a more reasonable SL. The LSAC was applied to a computer-aided detection (CAD) scheme to detect the initial polyp candidates, and experiments showed that it benefits the CAD scheme in both the detection sensitivity and specificity as compared to our previous work.

Virtual colonoscopy (VC) has gained popularity as a new colon diagnostic method over the last decade. VC is a new, less invasive alternative to the usually practiced optical colonoscopy for colorectal polyp and cancer screening, the second major cause of cancer related deaths in industrial nations. Haustral (colonic) folds serve as important landmarks for virtual endoscopic navigation in the existing computer-aided-diagnosis (CAD) system. In this paper, we propose and compare two different methods of haustral fold detection from volumetric computed tomographic virtual colonoscopy images. The colon lumen is segmented from the input using modified region growing and fuzzy connectedness. The first method for fold detection uses a level set that evolves on a mesh representation of the colon surface. The colon surface is obtained from the segmented colon lumen using the Marching Cubes algorithm. The second method for fold detection, based on a combination of heat diffusion and fuzzy c-means algorithm, is employed on the segmented colon volume. Folds obtained on the colon volume using this method are then transferred to the corresponding colon surface. After experimentation with different datasets, results are found to be promising. The results also demonstrate that the first method has a tendency of slight under-segmentation while the second method tends to slightly over-segment the folds.

Purpose: To compare the influence of visual and acoustical CAD markers on radiologist's performance with regard to suggestive and distractive effects. Materials and methods: Ten radiologists analyzed 150 pictures of chest CT slices. Every picture contained a visual CAD marker. 100 pictures showed one nodule: CAD marker marked this in 50 cases and in 50 cases a false positive finding (f.p.). The other 50 cases showed no nodule but an f.p. marker. After 3 years same images were presented to thirteen radiologists with only a sound as CAD marker. 55 of 150 images were marked, 30 true positive and 25 f.p. Sensitivity and f.p. rate were calculated for both marker types. Significance between sensitivities and f.p. rates were calculated by multiple-analysis-of-variance (MANOVA). Results: Without CAD mean sensitivity resp. f.p. were 57.7% /.13. In case of correct optical resp. acoustical marker sensitivity increased to 75.6% resp. 63.1%. For incorrect set marker mean f.p. rate increased to .31 resp. .24. MANOVA showed that marker's correctness highly significantly influenced sensitivity (p<.001) and f.n. (p=.005). Type of marker showed no significant influence on sensitivity (p=.26) or f.n. (p=.23) but on f.p. (p<.001). New work to be presented: Acoustical markers are a new means to increase radiologist's awareness of the presence of pulmonary nodules at CT scans with much less suggestive effect compared to optical markers. Conclusion: We found an unexpectedly low distraction effect for misplaced CAD markers. A suggestive effect was remarkable especially for optical markers. However acoustical markers offered less increase of sensitivity.

The growth rate of pulmonary nodules has been shown to be an indicator of malignancy, and previous work on pulmonary nodule characterization has suggested that the asymmetry of a nodule's shape may be correlated with malignancy. We have also observed that measurements in the axial direction on CT scans are less repeatable than measurements in-plane and this should be considered when making lesion size-change measurements. To address this, we present a method to measure the asymmetry of a pulmonary nodule's growth by the use of second-order central moments that are insensitive to z-direction variation. The difference in the moment ratios on each scan is used as a measure of the asymmetry of growth. To establish what level of difference is significant, the 95% confidence interval of the differences was determined on a zero-change dataset of 22 solid pulmonary nodules with repeat scans in the same session. This method was applied to a set of 47 solid, stable pulmonary nodules and a set of 49 solid, malignant nodules. The confidence interval established from the zero-change dataset was (-0.45, 0.38); nodules with differences outside this confidence interval are considered to have asymmetric growth. Of the 47 stable nodules, 12.8% (6/47) were found to have asymmetric growth compared to 24.5% (12/49) of malignant nodules. These preliminary results suggest that nodules with asymmetric growth can be identified.

Bronchial wall thickening is commonly observed in airway diseases. One method often used to quantitatively evaluate wall thickening in CT images is to estimate the ratio of the bronchial wall to the accompanying artery, or BWA ratio, and then assign a severity score based on the ratio. Assessment by visual inspection is unfortunately limited to airways perpendicular or parallel to the scanning plane. With high-resolution images from multi-detector CT scanners, it becomes possible to assess airways in any orientation. We selected CT scans from 20 patients with mild to severe COPD. A computer system automatically segmented each bronchial tree and measured the bronchial wall thicknesses. Next, neighboring arteries were detected and measured to determine BWA ratios. A score characterizing the extent and severity of wall thickening within each lobe was computed according to recommendations by Sheehan et al [1]. Two experienced radiologists independently scored wall thickening using visual assessment. Spearman's rank correlation showed a non-significant negative correlation (r=-0.1) between the computer and the reader average (p=0.4), while the correlation between readers was significant at r=0.65 (p=0.001). We subsequently identified 24 lobes with high discrepancies between visual and automated scoring. The readers re-examined those lobes and measured wall thickness using electronic calipers on perpendicular cross sections, rather than visual assessment. Using this more objective standard of wall thickness, the reader estimates of wall thickening increased to reach a significant positive correlation with automated scoring of r=0.65 (p=0.001). These results indicate that subjectivity is an important problem with visual evaluation, and that visual inspection may frequently underestimate disease extent and severity. Given that a manual evaluation of all airways is infeasible in routine clinical practice, we argue that automated methods should be developed and utilized.

Using computer-calculated features to characterize the shape of suspicious lesions aims to assist the diagnosis of pulmonary nodules; moreover, these computerized features have to be in agreement with radiologists' ratings measuring their human perception of the nodules' shape. In the Lung Image Database Consortium (LIDC), there exists strong disagreement among the radiologists on the ratings of the shape diagnostic characteristics as well as on their drawn outlines of the extent of the nodules. Since shape is often considered a property of the object boundary and the manual boundaries are not consistent among radiologists, new methods are necessary to, first, define regionbased boundaries that use radiologists' outlines as guides and, second, adapt computer-based shape measurements to use regions rather than the traditional nodule segmentation outlines. This paper introduces a method for defining a boundary region of interest by combining radiologist-drawn outlines (the pixel-set difference between the union and intersection of all radiologist-drawn outlines for a specific nodule), then adapts a radial gradient indexing method for use within image regions, and lastly predicts several composite ratings of sets of radiologists for shape-based characteristics: spiculation, lobulation, and sphericity. The prediction of the majority (mode) rating significantly outperforms earlier work on predicting the ratings of individual radiologists. The prediction of spiculation improves to 53% from 41%, lobulation increases to 44% from 38%, and sphericity improves to 58% from 43%. A binary version of the rating has high accuracy but poor Kappa agreement for all three shape characteristics.

A good problem representation is important for a pattern recognition system to be successful. The traditional approach to statistical pattern recognition is feature representation. More specifically, objects are represented by a number of features in a feature vector space, and classifiers are built in this representation. This is also the general trend in lung parenchyma classification in computed tomography (CT) images, where the features often are measures on feature histograms. Instead, we propose to build normal density based classifiers in dissimilarity representations for lung parenchyma classification. This allows for the classifiers to work on dissimilarities between objects, which might be a more natural way of representing lung parenchyma. In this context, dissimilarity is defined between CT regions of interest (ROI)s. ROIs are represented by their CT attenuation histogram and ROI dissimilarity is defined as a histogram dissimilarity measure between the attenuation histograms. In this setting, the full histograms are utilized according to the chosen histogram dissimilarity measure. We apply this idea to classification of different emphysema patterns as well as normal, healthy tissue. Two dissimilarity representation approaches as well as different histogram dissimilarity measures are considered. The approaches are evaluated on a set of 168 CT ROIs using normal density based classifiers all showing good performance. Compared to using histogram dissimilarity directly as distance in a emph{k} nearest neighbor classifier, which achieves a classification accuracy of $92.9%$, the best dissimilarity representation based classifier is significantly better with a classification accuracy of 97.0% ($text{emph{p" border="0" class="imgtopleft"> = 0.046$).

The chest x-ray radiological features of tuberculosis patients are well documented, and the radiological features that change in response to successful pharmaceutical therapy can be followed with longitudinal studies over time. The patients can also be classified as either responsive or resistant to pharmaceutical therapy based on clinical improvement. We have retrospectively collected time series chest x-ray images of 200 patients diagnosed with tuberculosis receiving the standard pharmaceutical treatment. Computer algorithms can be created to utilize image texture features to assess the temporal changes in the chest x-rays of the tuberculosis patients. This methodology provides a framework for a computer-assisted detection (CAD) system that may provide physicians with the ability to detect poor treatment response earlier in pharmaceutical therapy. Early detection allows physicians to respond with more timely treatment alternatives and improved outcomes. Such a system has the potential to increase treatment efficacy for millions of patients each year.

Multi-detector computed tomography (MDCT) has high accuracy and specificity on volumetrically capturing serial images of the lung. It increases the capability of computerized classification for lung tissue in medical research. This paper proposes a three-dimensional (3D) automated approach based on mathematical morphology and fuzzy logic for quantifying and classifying interstitial lung diseases (ILDs) and emphysema. The proposed methodology is composed of several stages: (1) an image multi-resolution decomposition scheme based on a 3D morphological filter is used to detect and analyze the different density patterns of the lung texture. Then, (2) for each pattern in the multi-resolution decomposition, six features are computed, for which fuzzy membership functions define a probability of association with a pathology class. Finally, (3) for each pathology class, the probabilities are combined up according to the weight assigned to each membership function and two threshold values are used to decide the final class of the pattern. The proposed approach was tested on 10 MDCT cases and the classification accuracy was: emphysema: 95%, fibrosis/honeycombing: 84% and ground glass: 97%.

Emphysema is a disease of the lungs that destroys the alveolar air sacs and induces long-term respiratory dysfunction. CT scans allow for the imaging of the anatomical basis of emphysema and quantification of the underlying disease state. Several measures have been introduced for the quantification emphysema directly from CT data; most,however, are based on the analysis of density information provided by the CT scans, which vary by scanner and can be hard to standardize across sites and time. Given that one of the anatomical variations associated with the progression of emphysema is the flatting of the diaphragm due to the loss of elasticity in the lung parenchyma, curvature analysis of the diaphragm would provide information about emphysema from CT. Therefore, we propose a new, non-density based measure of the curvature of the diaphragm that would allow for further quantification methods in a robust manner. To evaluate the new method, 24 whole-lung scans were analyzed using the ratios of the lung height and diaphragm width to diaphragm height as curvature estimates as well as using the emphysema index as comparison. Pearson correlation coefficients showed a strong trend of several of the proposed diaphragm curvature measures to have higher correlations, of up to r=0.57, with DLCO% and VA than did the emphysema index. Furthermore, we found emphysema index to have only a 0.27 correlation to the proposed measures, indicating that the proposed measures evaluate different aspects of the disease.

Advances in multi-detector technology have made CT pulmonary angiography (CTPA) a popular radiological tool for pulmonary emboli (PE) detection. CTPA provide rich detail of lung anatomy and is a useful diagnostic aid in highlighting even very small PE. However analyzing hundreds of slices is laborious and time-consuming for the practicing radiologist which may also cause misdiagnosis due to the presence of various PE look-alike. Computer-aided diagnosis (CAD) can be a potential second reader in providing key diagnostic information. Since PE occurs only in vessel arteries, it is important to mark this region of interest (ROI) during CAD preprocessing. In this paper, we present a new lung and vessel segmentation algorithm for extracting contrast-enhanced vessel ROI in CTPA. Existing approaches to segmentation either provide only the larger lung area without highlighting the vessels or is computationally prohibitive. In this paper, we propose a hybrid lung and vessel segmentation which uses an initial lung ROI and determines the vessels through a series of refinement steps. We first identify a coarse vessel ROI by finding the "holes" from the lung ROI. We then use the initial ROI as seed-points for a region-growing process while carefully excluding regions which are not relevant. The vessel segmentation mask covers 99% of the 259 PE from a real-world set of 107 CTPA. Further, our algorithm increases the net sensitivity of a prototype CAD system by 5-9% across all PE categories in the training and validation data sets. The average run-time of algorithm was only 100 seconds on a standard workstation.

The five year survival rate of the lung cancer is low with about twenty-five percent. In addition it is an obstinate lung cancer wherein three out of four people die within five years. Then, the early stage detection and treatment of the lung cancer are important. Recently, we can obtain CT and PET image at the same time because PET/CT device has been developed. PET/CT is possible for a highly accurate cancer diagnosis because it analyzes quantitative shape information from CT image and FDG distribution from PET image. However, neither benign-malignant classification nor staging intended for lung cancer have been established still enough by using PET/CT images. In this study, we detect lung nodules based on internal organs extracted from CT image, and we also develop algorithm which classifies benignmalignant and metastatic or non metastatic lung cancer using lung structure and FDG distribution(one and two hour after administering FDG). We apply the algorithm to 59 PET/CT images (malignant 43 cases [Ad:31, Sq:9, sm:3], benign 16 cases) and show the effectiveness of this algorithm.

The ability of a clinician to properly detect changes in the size of lung nodules over time is a vital element to both the diagnosis of malignant growths and the monitoring of the response of cancerous lesions to therapy. We have developed a novel metastasis sizing algorithm based on 3-D template matching with spherical tumor appearance models that were created to match the expected geometry of the tumors of interest while accounting for potential spatial offsets of nodules in the slice thickness direction. The spherical template that best-fits the overall volume of each lung metastasis was determined through the optimization of the 3-D normalized cross-correlation coefficients (NCCC) calculated between the templates and the nodules. A total of 17 different lung metastases were extracted manually from real patient CT datasets and reconstructed in 3-D using spherical harmonics equations to generate simulated nodules for testing our algorithm. Each metastasis 3-D shape was then subjected to 10%, 25%, 50%, 75% and 90% scaling of its volume to allow for 5 possible volume change combinations relative to the original size per each reconstructed nodule and inserted back into CT datasets with appropriate blurring and noise addition. When plotted against the true volume change, the nodule volume changes calculated by our algorithm for these 85 data points exhibited a high degree of accuracy (slope = 0.9817, R² = 0.9957). Our results demonstrate that the 3-D template matching method can be an effective, fast, and accurate tool for automated sizing of metastatic tumors.

An approach for optimum selection of lung nodule image characteristics in the feature domain is presented. This was applied to the classification module in the CAD system with data that was extracted from 42 ROI's of the 38 cases with an effective diameter of 3 to 8.5mm. 11 fundamental features were computed on the basis of dimensionality and image characteristics. The relation between the represented features of the 4 radiologists and the computed features was mapped using non-parametric correlation coefficients, multiple regression analysis and principle component analysis (PCA). Malignant and benign modules were classified based on the artificial neural network (ANN) to confirm the hypothesis from the mapping analysis. From the computed features and the radiologist's annotations, correlation coefficients between 0.2693 and 0.5178 were obtained. A combination of analyses namely regression, PCA, correlation and ANN were used to select optimum features. This resulted in F-test values of 0.821 and 0.643 for malignant and benign nodules respectively. The study of the relationship between the features and the weightage towards each of the representative classes resulted in optimum feature input for a CAD system. A composite analysis derived from correlation, PCA, multiple regression and the classification algorithm, collectively termed as the knowledge base, was used arrive at an "optimum" set of lung nodule features.

CT images are considered as effective for differential diagnosis of diffuse lung diseases. However, the diagnosis of diffuse lung diseases is a difficult problem for the radiologists, because they show a variety of patterns on CT images. So, our purpose is to construct a computer-aided diagnosis (CAD) system for classification of patterns for diffuse lung diseases in thoracic CT images, which gives both quantitative and objective information as a second opinion, to decrease the burdens of radiologists. In this article, we propose a CAD system based on the conventional pattern recognition framework, which consists of two sub-systems; one is feature extraction part and the other is classification part. In the feature extraction part, we adopted a Gabor filter, which can extract patterns such like local edges and segments from input textures, as a feature extraction of CT images. In the recognition part, we used a boosting method. Boosting is a kind of voting method by several classifiers to improve decision precision. We applied AdaBoost algorithm for boosting method. At first, we evaluated each boosting component classifier, and we confirmed they had not enough performances in classification of patterns for diffuse lung diseases. Next, we evaluated the performance of boosting method. As a result, by use of our system, we could improve the classification rate of patterns for diffuse lung diseases.

The successful development of high performance computer-aided-diagnostic systems has potential to assist radiologists in the detection and diagnosis of diffuse lung disease. We developed in this study an automated scheme for the detection of diffuse lung disease on multi-detector computed tomography (MDCT). Our database consisted of 68 CT scans, which included 31 normal and 37 abnormal cases with three kinds of abnormal patterns, i.e., ground glass opacity, reticular, and honeycombing. Two radiologists first selected the CT scans with abnormal patterns based on clinical reports. The areas that included specific abnormal patterns in the selected CT images were then delineated as reference standards by an expert chest radiologist. To detect abnormal cases with diffuse lung disease, the lungs were first segmented from the background in each slice by use of a texture analysis technique, and then divided into contiguous volumes of interest (VOIs) with a 64×64×64 matrix size. For each VOI, we calculated many statistical texture features, including the mean and standard deviation of CT values, features determined from the run length matrix, and features from the co-occurrence matrix. A quadratic classifier was employed for distinguishing between normal and abnormal VOIs by use of a leave-one-case-out validation scheme. A rule-based criterion was employed to further determine whether a case was normal or abnormal. For the detection of abnormal VOIs, our CAD system achieved a sensitivity of 86% and a specificity of 90%. For the detection of abnormal cases, it achieved a sensitivity of 89% and a specificity of 90%. This preliminary study indicates that our CAD system would be useful for the detection of diffuse lung disease.

In this paper, a cascade classification scheme was proposed to improve computational efficiency in lung parenchyma quantification in HRCT images. Proposed cascade classification scheme includes four steps: cost-based class-specific feature selection, class-specific classifier training, classifier ordering, cascade feature extraction and classification. In the first step, feature sets were determined by sequential forward floating selection (SFFS) using performance improvement to extraction cost ratio criterion. Then classifiers were trained to classify specific class from all of other classes. Using accuracies of those classifiers, the order of classification was determined; from the highest accuracy to lowest accuracy. To quantify new images, feature extraction and classification were sequentially repeated. The impact of using the proposed cascade classification scheme is evaluated in terms of computational cost and classification accuracy. For automated classification, support vector machine (SVM) was implemented. To assess the performance and crossvalidation of the system, ten-folding method was used. In the experimental results, the computational cost was reduced by 46% and the overall accuracy was 92.04% which is not significantly different in a comparison of conventional method. This work shows that, in our classification problem, using the proposed cascade classification scheme can reduce the computational cost in the feature extraction while maintaining the classification accuracy.

Chronic Obstructive Pulmonary Disease (COPD) is often characterized by partial or complete obstruction of airflow in the lungs. This can be due to airway wall thickening and retained secretions, resulting in foci of mucoid impactions. Although radiologists have proposed scoring systems to assess extent and severity of airway diseases from CT images, these scores are seldom used clinically due to impracticality. The high level of subjectivity from visual inspection and the sheer number of airways in the lungs mean that automation is critical in order to realize accurate scoring. In this work we assess the feasibility of including an automated mucus detection method in a clinical scoring system. Twenty high-resolution datasets of patients with mild to severe bronchiectasis were randomly selected, and used to test the ability of the computer to detect the presence or absence of mucus in each lobe (100 lobes in all). Two experienced radiologists independently scored the presence or absence of mucus in each lobe based on the visual assessment method recommended by Sheehan et al [1]. These results were compared with an automated method developed for mucus plug detection [2]. Results showed agreement between the two readers on 44% of the lobes for presence of mucus, 39% of lobes for absence of mucus, and discordant opinions on 17 lobes. For 61 lobes where 1 or both readers detected mucus, the computer sensitivity was 75.4%, the specificity was 69.2%, and the positive predictive value (PPV) was 79.3%. Six computer false positives were a-posteriori reviewed by the experts and reassessed as true positives, yielding results of 77.6% sensitivity, 81.8% for specificity, and 89.6% PPV.

One of the fundamental issues in bridging the gap between the proliferation of Content-Based Image Retrieval (CBIR) systems in the scientific literature and the deficiency of their usage in medical community is based on the characteristic of CBIR to access information by images or/and text only. Yet, the way physicians are reasoning about patients leads intuitively to a case representation. Hence, a proper solution to overcome this gap is to consider a CBIR approach inspired by Case-Based Reasoning (CBR), which naturally introduces medical knowledge structured by cases. Moreover, in a CBR system, the knowledge is incrementally added and learned. The purpose of this study is to initiate a translational solution from CBIR algorithms to clinical practice, using a CBIR/CBR hybrid approach. Therefore, we advance the idea of a translational incremental similarity-based reasoning (TISBR), using combined CBIR and CBR characteristics: incremental learning of medical knowledge, medical case-based structure of the knowledge (CBR), image usage to retrieve similar cases (CBIR), similarity concept (central for both paradigms). For this purpose, three major axes are explored: the indexing, the cases retrieval and the search refinement, applied to Breast Cancer Grading (BCG), a powerful breast cancer prognosis exam. The effectiveness of this strategy is currently evaluated over cases provided by the Pathology Department of Singapore National University Hospital, for the indexing. With its current accuracy, TISBR launches interesting perspectives for complex reasoning in future medical research, opening the way to a better knowledge traceability and a better acceptance rate of computer-aided diagnosis assistance among practitioners.

Focusing is a critical step in microscope observation of slides. Autoscanning microscopes have to perform autofocus function accurately to record high quality images that may be later analyzed using sophisticated algorithms. Video based autofocus has become a viable option due to the availability of high computing power and cameras that provide high resolution images. A focus function which obtains a peak value when an image in focus has been encountered is used by these methods. In this paper a novel focus function based on the shape of the objects being observed is proposed. A segmentation based approach to autofocus blood smears where the primary objects being observed, the red blood cells (RBC), are circular is presented. The scheme first segments the RBCs and then determines the valid RBCs by using area criterion. The average form factor and eccentricity values of the valid RBCs in a given frame are computed. A plot of these parameters vs. the frame number will result in a peak and trough respectively for the in focus image. Results presented for various data sets show that form factor is a suitable autofocus function.

Pathologists examine histology sections to make diagnostic and prognostic assessments regarding cancer based on deviations in cellular and/or glandular structures. However, these assessments are subjective and exhibit some degree of observer variability. Recent studies have shown that fractal dimension (a quantitative measure of structural complexity) has proven useful for characterizing structural deviations and exhibits great potential for automated cancer diagnosis and prognosis. Computing fractal dimension relies on accurate image segmentation to capture the architectural complexity of the histology specimen. For this purpose, previous studies have used techniques such as intensity histogram analysis and edge detection algorithms. However, care must be taken when segmenting pathologically relevant structures since improper edge detection can result in an inaccurate estimation of fractal dimension. In this study, we established a reliable method for segmenting edges from grayscale images. We used a Koch snowflake, an object of known fractal dimension, to investigate the accuracy of various edge detection algorithms and selected the most appropriate algorithm to extract the outline structures. Next, we created validation objects ranging in fractal dimension from 1.3 to 1.9 imitating the size, structural complexity, and spatial pixel intensity distribution of stained histology section images. We applied increasing intensity thresholds to the validation objects to extract the outline structures and observe the effects on the corresponding segmentation and fractal dimension. The intensity threshold yielding the maximum fractal dimension provided the most accurate fractal dimension and segmentation, indicating that this quantitative method could be used in an automated classification system for histology specimens.

Distance metrics are often used as a way to compare the similarity of two objects, each represented by a set of features in high-dimensional space. The Euclidean metric is a popular distance metric, employed for a variety of applications. Non-Euclidean distance metrics have also been proposed, and the choice of distance metric for any specific application or domain is a non-trivial task. Furthermore, most distance metrics treat each dimension or object feature as having the same relative importance in determining object similarity. In many applications, such as in Content-Based Image Retrieval (CBIR), where images are quantified and then compared according to their image content, it may be beneficial to utilize a similarity metric where features are weighted according to their ability to distinguish between object classes. In the CBIR paradigm, every image is represented as a vector of quantitative feature values derived from the image content, and a similarity measure is applied to determine which of the database images is most similar to the query. In this work, we present a boosted distance metric (BDM), where individual features are weighted according to their discriminatory power, and compare the performance of this metric to 9 other traditional distance metrics in a CBIR system for digital histopathology. We apply our system to three different breast tissue histology cohorts - (1) 54 breast histology studies corresponding to benign and cancerous images, (2) 36 breast cancer studies corresponding to low and high Bloom-Richardson (BR) grades, and (3) 41 breast cancer studies with high and low levels of lymphocytic infiltration. Over all 3 data cohorts, the BDM performs better compared to 9 traditional metrics, with a greater area under the precision-recall curve. In addition, we performed SVM classification using the BDM along with the traditional metrics, and found that the boosted metric achieves a higher classification accuracy (over 96%) in distinguishing between the tissue classes in each of 3 data cohorts considered. The 10 different similarity metrics were also used to generate similarity matrices between all samples in each of the 3 cohorts. For each cohort, each of the 10 similarity matrices were subjected to normalized cuts, resulting in a reduced dimensional representation of the data samples. The BDM resulted in the best discrimination between tissue classes in the reduced embedding space.

This paper introduces an automated method for finding diagnostic regions-of-interest (RoIs) in histopathological images. This method is based on the cognitive process of visual selective attention that arises during a pathologist's image examination. Specifically, it emulates the first examination phase, which consists in a coarse search for tissue structures at a "low zoom" to separate the image into relevant regions.1 The pathologist's cognitive performance depends on inherent image visual cues - bottom-up information - and on acquired clinical medicine knowledge - top-down mechanisms -. Our pathologist's visual attention model integrates the latter two components. The selected bottom-up information includes local low level features such as intensity, color, orientation and texture information. Top-down information is related to the anatomical and pathological structures known by the expert. A coarse approximation to these structures is achieved by an oversegmentation algorithm, inspired by psychological grouping theories. The algorithm parameters are learned from an expert pathologist's segmentation. Top-down and bottom-up integration is achieved by calculating a unique index for each of the low level characteristics inside the region. Relevancy is estimated as a simple average of these indexes. Finally, a binary decision rule defines whether or not a region is interesting. The method was evaluated on a set of 49 images using a perceptually-weighted evaluation criterion, finding a quality gain of 3dB when comparing to a classical bottom-up model of attention.

Prostate cancer is the most common type of cancer and the second leading cause of cancer death among men in US1. Quantitative assessment of prostate histology provides potential automatic classification of prostate lesions and prediction of response to therapy. Traditionally, prostate cancer diagnosis is made by the analysis of prostate-specific antigen (PSA) levels and histopathological images of biopsy samples under microscopes. In this application, we utilize a texture analysis method based on the run-length matrix for identifying tissue abnormalities in prostate histology. A tissue sample was collected from a radical prostatectomy, H&E fixed, and assessed by a pathologist as normal tissue or prostatic carcinoma (PCa). The sample was then subsequently digitized at 50X magnification. We divided the digitized image into sub-regions of 20 X 20 pixels and classified each sub-region as normal or PCa by a texture analysis method. In the texture analysis, we computed texture features for each of the sub-regions based on the Gray-level Run-length Matrix(GL-RLM). Those features include LGRE, HGRE and RPC from the run-length matrix, mean and standard deviation of the pixel intensity. We utilized a feature selection algorithm to select a set of effective features and used a multi-layer perceptron (MLP) classifier to distinguish normal from PCa. In total, the whole histological image was divided into 42 PCa and 6280 normal regions. Three-fold cross validation results show that the proposed method achieves an average classification accuracy of 89.5% with a sensitivity and specificity of 90.48% and 89.49%, respectively.

Screening and detection of prostate cancer (CaP) currently lacks an image-based protocol which is reflected in the high false negative rates currently associated with blinded sextant biopsies. Multi-protocol magnetic resonance imaging (MRI) offers high resolution functional and structural data about internal body structures (such as the prostate). In this paper we present a novel comprehensive computer-aided scheme for CaP detection from high resolution in vivo multi-protocol MRI by integrating functional and structural information obtained via dynamic-contrast enhanced (DCE) and T2-weighted (T2-w) MRI, respectively. Our scheme is fully-automated and comprises (a) prostate segmentation, (b) multimodal image registration, and (c) data representation and multi-classifier modules for information fusion. Following prostate boundary segmentation via an improved active shape model, the DCE/T2-w protocols and the T2-w/ex vivo histological prostatectomy specimens are brought into alignment via a deformable, multi-attribute registration scheme. T2-w/histology alignment allows for the mapping of true CaP extent onto the in vivo MRI, which is used for training and evaluation of a multi-protocol MRI CaP classifier. The meta-classifier used is a random forest constructed by bagging multiple decision tree classifiers, each trained individually on T2-w structural, textural and DCE functional attributes. 3-fold classifier cross validation was performed using a set of 18 images derived from 6 patient datasets on a per-pixel basis. Our results show that the results of CaP detection obtained from integration of T2-w structural textural data and DCE functional data (area under the ROC curve of 0.815) significantly outperforms detection based on either of the individual modalities (0.704 (T2-w) and 0.682 (DCE)). It was also found that a meta-classifier trained directly on integrated T2-w and DCE data (data-level integration) significantly outperformed a decision-level meta-classifier, constructed by combining the classifier outputs from the individual T2-w and DCE channels.

The aim of this research was to investigate the performance of wavelet transform based features of ultrasound radiofrequency (RF) time series for automated detection of prostate cancer tumors in transrectal ultrasound images. Sequential frames of RF echo signals from 35 extracted prostate specimens were recorded in parallel planes, while the ultrasound probe and the tissue were fixed in position in each imaging plane. The sequence of RF echo signal samples corresponding to a particular spot in tissue imaging plane constitutes one RF time series. Each region of interest (ROI) of ultrasound image was represented by three groups of features of its time series, namely, wavelet, spectral and fractal features. Wavelet transform approximation and detail sequences of each ROI were averaged and used as wavelet features. The average value of the normalized spectrum in four quarters of the frequency range along with the intercept and slope of a regression line fitted to the values of the spectrum versus normalized frequency plot formed six spectral features. Fractal dimension (FD) of the RF time series were computed based on the Higuchi's approach. A support vector machine (SVM) classifier was used to classify the ROIs. The results indicate that combining wavelet coefficient based features with previously proposed spectral and fractal features of RF time series data would increase the area under ROC curve from 93.1% to 95.0%, respectively. Furthermore, the accuracy, sensitivity, and specificity increases to 91.7%, 86.6%, and 94.7%, from 85.7%, 85.2%, and 86.1%, respectively, using only spectral and fractal features.

Glaucoma is an irreversible ocular disease leading to permanent blindness. However, early detection can be effective in slowing or halting the progression of the disease. Physiologically, glaucoma progression is quantified by increased excavation of the optic cup. This progression can be quantified in retinal fundus images via the optic cup to disc ratio (CDR), since in increased glaucomatous neuropathy, the relative size of the optic cup to the optic disc is increased. The ARGALI framework constitutes of various segmentation approaches employing level set, color intensity thresholds and ellipse fitting for the extraction of the optic cup and disc from retinal images as preliminary steps. Following this, different combinations of the obtained results are then utilized to calculate the corresponding CDR values. The individual results are subsequently fused using a neural network. The learning function of the neural network is trained with a set of 100 retinal images For testing, a separate set 40 images is then used to compare the obtained CDR against a clinically graded CDR, and it is shown that the neural network-based result performs better than the individual components, with 96% of the results within intra-observer variability. The results indicate good promise for the further development of ARGALI as a tool for the early detection of glaucoma.

A large cup-to-disc (C/D) ratio, which is the ratio of the diameter of the depression (cup) to that of the optical nerve head (ONH, disc), can be one of the important signs for diagnosis of glaucoma. Eighty eyes, including 25 eyes with the signs of glaucoma, were imaged by a stereo retinal fundus camera. An ophthalmologist provided the outlines of cup and disc on a regular monitor and on the stereo display. The depth image of the ONH was created by determining the corresponding pixels in a pair of images based on the correlation coefficient in localized regions. The areas of the disc and cup were determined by use of the red component in one of the color images and by use of the depth image, respectively. The C/D ratio was determined based on the largest vertical lengths in the cup and disc areas, which was then compared with that by the ophthalmologist. The disc areas determined by the computerized method agreed relatively well with those determined by the ophthalmologist, whereas the agreement for the cup areas was somewhat lower. When C/D ratios were employed for distinction between the glaucomatous and non-glaucomatous eyes, the area under the receiver operating characteristic curve (AUC) was 0.83. The computerized analysis of ONH can be useful for diagnosis of glaucoma.

We hypothesize that the vertebra-to-vertebra patterns of spinal flexion and extension motion of persons with lower back pain will differ from those of persons who are pain-free. Thus, it is our goal to measure the motion of individual lumbar vertebrae noninvasively from dynamic fluoroscopic sequences. Two-dimensional normalized mutual information-based image registration was used to track frame-to-frame motion. Software was developed that required the operator to identify each vertebra on the first frame of the sequence using a four-point "caliper" placed at the posterior and anterior edges of the inferior and superior end plates of the target vertebrae. The program then resolved the individual motions of each vertebra independently throughout the entire sequence. To validate the technique, 6 cadaveric lumbar spine specimens were potted in polymethylmethacrylate and instrumented with optoelectric sensors. The specimens were then placed in a custom dynamic spine simulator and moved through flexion-extension cycles while kinematic data and fluoroscopic sequences were simultaneously acquired. We found strong correlation between the absolute flexionextension range of motion of each vertebra as recorded by the optoelectric system and as determined from the fluoroscopic sequence via registration. We conclude that this method is a viable way of noninvasively assessing twodimensional vertebral motion.

Developmental dysplasia of the hip is a congenital hip joint malformation affecting the proximal femurs and acetabulum that are subluxatable, dislocatable, and dislocated. Conventionally, physicians made diagnoses and treatments only based on findings from two-dimensional (2D) images by manually calculating clinic parameters. However, anatomical complexity of the disease and the limitation of current standard procedures make accurate diagnosis quite difficultly. In this study, we developed a system that provides quantitative measurement of 3D clinical indexes based on computed tomography (CT) images. To extract bone structure from surrounding tissues more accurately, the system firstly segments the bone using a knowledge-based fuzzy clustering method, which is formulated by modifying the objective function of the standard fuzzy c-means algorithm with additive adaptation penalty. The second part of the system calculates automatically the clinical indexes, which are extended from 2D to 3D for accurate description of spatial relationship between femurs and acetabulum. To evaluate the system performance, experimental study based on 22 patients with unilateral or bilateral affected hip was performed. The results of 3D acetabulum index (AI) automatically provided by the system were validated by comparison with 2D results measured by surgeons manually. The correlation between the two results was found to be 0.622 (p<0.01).

A novel method is proposed for the automatic detection of fetal spine in ultrasound images along with its orientation in this paper. This problem presents a variety of challenges, including robustness to speckle noise, variations in the visible shape of the spine due to orientation of the ultrasound probe with respect to the fetus and the lack of a proper edge enclosing the entire spine on account of its composition out of distinct vertebra. The proposed method improves robustness and accuracy by making use of two independent techniques to estimate the spine, and then detects the exact location using a cross-correlation approach. Experimental results show that the proposed method is promising for fetal spine detection.

We present a new method for automatic detection of the lumbar vertebrae and disk structure from MR images. In clinical settings, radiologists utilize several images of the lumbar structure for diagnosis of lumbar disorders. These images are co-registered by technicians and represent orthogonal features of the lumbar region. We combine information from T1W sagittal, T2W sagittal and T2W axial MR images to automatically label disks and vertebral columns. The method couples geometric and tissue property information available from the three types of images with image analysis approaches to achieve 98.8% accuracy for the disk labeling task on a test set of 67 images containing 335 disks.

We present two novel metrics for assessing scoliosis, in which the geometric centers of all the affected vertebrae in an antero-posterior (A-P) radiographic image are used. This is in contradistinction to the existing methods of using selected vertebrae, and determining either their endplates or the intersections of their diagonals, to define a scoliotic angle. Our first metric delivers a scoliotic angle, comparable to the Cobb and Ferguson angles. It measures the sum of the angles between the centers of the affected vertebrae, and avoids the need for an observer to decide on the extent of component curvatures. Our second metric calculates the normalized root-mean-square curvature of the smoothest path comprising piece-wise polynomial splines fitted to the geometric centers of the vertebrae. The smoothest path is useful in modeling the spinal curvature. Our metrics were compared to existing methods using radiographs from a group of twenty subjects with spinal curvatures of varying severity. Their values were strongly correlated with those of the scoliotic angles (r = 0.850 - 0.886), indicating that they are valid surrogates for measuring the severity of scoliosis. Our direct use of positional data removes the vagaries of determining variably shaped endplates, and circumvented the significant interand intra-observer errors of the Cobb and Ferguson methods. Although we applied our metrics to two-dimensional (2- D) data in this paper, they are equally applicable to three-dimensional (3-D) data. We anticipate that they will prove to be the basis for a reliable 3-D measurement and classification system.

We are developing a CAD system for automated bladder segmentation and detection of bladder lesions on MDCT urography, which potentially can assist radiologists in detecting bladder cancer. In the first stage of our CAD system, given a starting point, the bladder is segmented based on 3D region growing and active contours. In the second stage, lesion candidates are detected using histogram and shape analysis to separate the abnormality from the background, which is the bladder partially filled with contrast material. In this pilot study, a limited data set of 15 patients with 29 biopsy-proven lesions (26 malignant, 3 benign) was used. The average size for the 26 malignant lesions was 10 mm (range: 4.2 mm - 30.5mm) with conspicuity in the range of 2 to 5 on a 5-point scale (5=very subtle). The average size for the 3 benign lesions was 14 mm (range: 3.5 mm - 25mm) with conspicuity in the range of 2 to 3. Our segmentation program successfully segmented both the contrast and non-contrast part of the bladder in 87% (13/15) of the patients. The contrast-filled bladder region was successfully segmented for all 15 patients. Our system detected 83% (24/29) of the lesions with 1.4 (21/15) false positives per patient. 85% (22/26) of the bladder cancers were detected. The main cause for missed lesions was that they were in the non-contrast bladder region, which was not included in the detection stage in this pilot study. The results demonstrate the feasibility of developing a CAD system for automated segmentation of the bladder and detection of bladder malignancies.

Cancer of the esophagus has the worst prediction of all known cancers in Germany. The early detection of suspicious changes in the esophagus allows therapies that can prevent the cancer. Barrett's esophagus is a premalignant change of the esophagus that is a strong indication for cancer. Therefore there is a big interest to detect Barrett's esophagus as early as possible. The standard examination is done with a videoscope where the physician checks the esophagus for suspicious regions. Once a suspicious region is found, the physician takes a biopsy of that region to get a histological result of it. Besides the traditional white light for the illumination there is a new technology: the so called narrow-band Imaging (NBI). This technology uses a smaller spectrum of the visible light to highlight the scene captured by the videoscope. Medical studies indicate that the use of NBI instead of white light can increase the rate of correct diagnoses of a physician. In the future, Computer-Assisted Diagnosis (CAD) which is well known in the area of mammography might be used to support the physician in the diagnosis of different lesions in the esophagus. A knowledge-based system which uses a database is a possible solution for this task. For our work we have collected NBI images containing 326 Regions of Interest (ROI) of three typical classes: epithelium, cardia mucosa and Barrett's esophagus. We then used standard texture analysis features like those proposed by Haralick, Chen, Gabor and Unser to extract features from every ROI. The performance of the classification was evaluated with a classifier using the leaving-one-out sampling. The best result that was achieved is an accuracy of 92% for all classes and an accuracy of 76% for Barrett's esophagus. These results show that the NBI technology can provide a good diagnosis support when used in a CAD system.

We present a method for patient-adaptive detection of bleeding region for a Capsule Endoscopy (CE) images. The CE system has 320x320 resolution and transmits 3 images per second to receiver during around 10-hour. We have developed a technique to detect the bleeding automatically utilizing color spectrum transformation (CST) method. However, because of irregular conditions like organ difference, patient difference and illumination condition, detection performance is not uniform. To solve this problem, the detection method in this paper include parameter compensation step which compensate irregular image condition using color balance index (CBI). We have investigated color balance through sequential 2 millions images. Based on this pre-experimental result, we defined ΔCBI to represent deviate of color balance compared with standard small bowel color balance. The ΔCBI feature value is extracted from each image and used in CST method as parameter compensation constant. After candidate pixels were detected using CST method, they were labeled and examined with a bleeding character. We tested our method with 4,800 images in 12 patient data set (9 abnormal, 3 normal). Our experimental results show the proposed method achieves (before patient adaptive method : 80.87% and 74.25%, after patient adaptive method : 94.87% and 96.12%) of sensitivity and specificity.

Computer-aided diagnosis (CAD) systems usually require information about regions of interest in images, like: lungs (for nodule detection), colon (for identifying polyps), etc. Many times, it is computationally intensive to process large data sets as in CT to find these areas of interest. A fast and accurate recognition of the different regions of interest in the human body from images is therefore necessary. In this paper we propose a fast and efficient algorithm that can detect the organs of interest in a CT volume and estimate their sizes. Instead of analyzing the whole 3D volume; which is computationally expensive, a binary search technique is adapted to search in a few slices. The slices selected in the search process is segmented and different regions are labeled. Decision over whether the image belongs to a lung or colon or both is based on the count of lung/colon pixels in the slice. Once the detection is done we look for the start and end slice of the body part to have an estimate of their sizes. The algorithm involves certain search decisions based on some predefined threshold values which are empirically chosen from a training data set. The effectiveness of our technique is confirmed by applying it on an independent test data set. Detection accuracy of 100% is obtained on a test set. This algorithm can be integrated in a CAD system for running the right application, or can be used in pre-sets for visualization purposes and other post-processing like image registration etc.

Knowledge of the exact shape of a lesion, or ground truth (GT), is necessary for algorithm validation, measurement metric analysis, accurate size estimation. When multiple readers provide their documentations of a lesion that can ultimately be described with occupancy regions, estimating the unknown GT is achieved by aptly merging those occupancy regions into a single outcome. Several methods are already available but, even when they consider the spatial location of pixels, e.g. thresholded probability-map (TPM) or STAPLE, pixels are assumed spatially independent (even when STAPLE proposes a hidden-Markov-random-field fix). In this paper we propose Truth Estimate from Self Distances (TESD): a new voting scheme, for all the voxels inside and outside the occupancy region, in order to take in account three key characteristics: (a) critical shape conformations, like holes or spikes, that are defined by the reciprocally surrounding pixels, (b) marking co-locations, meaning the closeness without intersection of one reader's marking to other readers' ones and c) the three-dimensionality of lesions as imaged by CT scanners. In TESD each voxel is labeled into four categories according to its signed distance transform and then the labeled images are combined with a center of gravity method to provide the GT estimation. This theoretical approach was validated on a subset of the publicly available Lung Image Database Consortium archive, where a total of 35 nodules documented on 26 scans by all four radiologists were available. The results obtained are reasonable estimates, with GT obtained close to TPM and STAPLE; at the same time this method is not limited to the intersections of readers' marked regions.

Computer-aided diagnosis usually screens thousands of instances to find only a few positive cases that indicate probable presence of disease.The amount of patient data increases consistently all the time. In diagnosis of new instances, disagreement occurs between a CAD system and physicians, which suggests inaccurate classifiers. Intuitively, misclassified instances and the previously acquired data should be used to retrain the classifier. This, however, is very time consuming and, in some cases where dataset is too large, becomes infeasible. In addition, among the patient data, only a small percentile shows positive sign, which is known as imbalanced data.We present an incremental Support Vector Machines(SVM) as a solution for the class imbalance problem in classification of anomaly in medical images. The support vectors provide a concise representation of the distribution of the training data. Here we use bootstrapping to identify potential candidate support vectors for future iterations. Experiments were conducted using images from endoscopy videos, and the sensitivity and specificity were close to that of SVM trained using all samples available at a given incremental step with significantly improved efficiency in training the classifier.

We present an automatic method to quickly and accurately detect multiple anatomy region-of-interests (ROIs) from CT topogram images. Our method first detects a redundant and potentially erroneous set of local features. Their spatial configurations are captured by a set of local voting functions. Unlike all the existing methods where the idea was to try to "hit" the correct/best constellations of local features, we have taken an opposite approach. We try to peel away the bad features until a safe (i.e., conservatively small) number of features remain. It is deterministic in nature and guarantees a success even for extremely noisy cases. The advantages of the method are its robustness and computational efficiency. Our method also addresses the potential scenario in which outliers (i.e., false landmarks detections) forms plausible configurations. As long as such outliers are a minority, the method can successfully remove these outliers. The final ROI of the anatomy is computed from a best subset of the remaining local features. Experimental validation was carried out for multiple organs detection from a large collection of CT topogram images. Fast and highly robust performance was observed. In the testing data sets, the detection rate varies from 98.2% to 100% for different ROIs and the false detection rate is from 0.0% to 0.5% for different ROIs. The method is fast and accurate enough to be seamlessly integrated into a real-time work flow on the CT machine to improve efficiency, consistency, and repeatability.