This PDF file contains the front matter associated with SPIE Proceedings Volume 8315, including the Title Page, Copyright information, Table of Contents, Introduction, and the Conference Committee listing.

An automated cell nuclei detection algorithm is described to be used for the quantification of immunohistochemicallystained tissues. Detection and segmentation of positively stained cells and their separation from the background and negatively-stained cells is crucial for fast, accurate, consistent and objective analysis of pathology images. One of the major challenges is the identification, hence accurate counting of individual cells, when these cells form clusters. To identify individual cell nuclei within clusters, we propose a new cell nuclei detection method based on the well-known watershed segmentation, which can lead to under- or over-segmentation for this problem. Our algorithm handles oversegmentation by combining H-minima transformed watershed algorithm with a novel region merging technique. To handle under-segmentation problem, we develop a Laplacian-of-Gaussian (LoG) filtering based blob detection algorithm, which estimates the range of the scales from the image adaptively. An SVM classifier was trained in order to separate non-touching single cells and touching cell clusters with five features representing connected region properties such as eccentricity, area, perimeter, convex area and perimeter-to-area ratio. Classified touching cell clusters are segmented with the H-minima based watershed algorithm. The resulting over-segmented regions are improved with the merging algorithm. The remaining under-segmented cell clusters are convolved with LoG filters to detect the cells within them. Cell-by-cell nucleus detection performance is evaluated by comparing computer detections with cell locations manually marked by eight pathology residents. The sensitivity is 89% when the cells are marked as positive at least by one resident and it increases to 99% when the evaluated cells are marked by all eight residents. In comparison, the average reader sensitivity varies between 70% ± 18% and 95% ± 11%.

We present a computationally efficient method for analyzing H&E stained digital pathology slides with the objective of discriminating diagnostically relevant vs. irrelevant regions. Such technology is useful for several applications: (1) It can speed up computer aided diagnosis (CAD) for histopathology based cancer detection and grading by an order of magnitude through a triage-like preprocessing and pruning. (2) It can improve the response time for an interactive digital pathology workstation (which is usually dealing with several GByte digital pathology slides), e.g., through controlling adaptive compression or prioritization algorithms. (3) It can support the detection and grading workflow for expert pathologists in a semi-automated diagnosis, hereby increasing throughput and accuracy. At the core of the presented method is the statistical characterization of tissue components that are indicative for the pathologist's decision about malignancy vs. benignity, such as, nuclei, tubules, cytoplasm, etc. In order to allow for effective yet computationally efficient processing, we propose visual descriptors that capture the distribution of color intensities observed for nuclei and cytoplasm. Discrimination between statistics of relevant vs. irrelevant regions is learned from annotated data, and inference is performed via linear classification. We validate the proposed method both qualitatively and quantitatively. Experiments show a cross validation error rate of 1.4%. We further show that the proposed method can prune ≈90% of the area of pathological slides while maintaining 100% of all relevant information, which allows for a speedup of a factor of 10 for CAD systems.

Automated 3D breast ultrasound (ABUS) is a novel imaging modality, in which motorized scans of the breasts are made with a wide transducer through a membrane under modest compression. The technology has gained high interest and may become widely used in screening of dense breasts, where sensitivity of mammography is poor. ABUS has a high sensitivity for detecting solid breast lesions. However, reading ABUS images is time consuming, and subtle abnormalities may be missed. Therefore, we are developing a computer aided detection (CAD) system to help reduce reading time and errors. In the multi-stage system we propose, segmentations of the breast and nipple are performed, providing landmarks for the detection algorithm. Subsequently, voxel features characterizing coronal spiculation patterns, blobness, contrast, and locations with respect to landmarks are extracted. Using an ensemble of classifiers, a likelihood map indicating potential malignancies is computed. Local maxima in the likelihood map are determined using a local maxima detector and form a set of candidate lesions in each view. These candidates are further processed in a second detection stage, which includes region segmentation, feature extraction and a final classification. Region segmentation is performed using a 3D spiral-scanning dynamic programming method. Region features include descriptors of shape, acoustic behavior and texture. Performance was determined using a 78-patient dataset with 93 images, including 50 malignant lesions. We used 10-fold cross-validation. Using FROC analysis we found that the system obtains a lesion sensitivity of 60% and 70% at 2 and 4 false positives per image respectively.

Feature extraction is a critical component of medical image analysis. Many computer-aided diagnosis approaches employ hand-designed, heuristic lesion extracted features. An alternative approach is to learn features directly from images. In this preliminary study, we explored the use of Adaptive Deconvolutional Networks (ADN) for learning high-level features in diagnostic breast mass lesion images with potential application to computer-aided diagnosis (CADx) and content-based image retrieval (CBIR). ADNs (Zeiler, et. al., 2011), are recently-proposed unsupervised, generative hierarchical models that decompose images via convolution sparse coding and max pooling. We trained the ADNs to learn multiple layers of representation for two breast image data sets on two different modalities (739 full field digital mammography (FFDM) and 2393 ultrasound images). Feature map calculations were accelerated by use of GPUs. Following Zeiler et. al., we applied the Spatial Pyramid Matching (SPM) kernel (Lazebnik, et. al., 2006) on the inferred feature maps and combined this with a linear support vector machine (SVM) classifier for the task of binary classification between cancer and non-cancer breast mass lesions. Non-linear, local structure preserving dimension reduction, Elastic Embedding (Carreira-Perpiñán, 2010), was then used to visualize the SPM kernel output in 2D and qualitatively inspect image relationships learned. Performance was found to be competitive with current CADx schemes that use human-designed features, e.g., achieving a 0.632+ bootstrap AUC (by case) of 0.83 [0.78, 0.89] for an ultrasound image set (1125 cases).

Breast MRI has emerged as an effective modality for the clinical management of breast cancer. Evidence suggests that computer-aided applications can further improve the diagnostic accuracy of breast MRI. A critical and challenging first step for automated breast MRI analysis, is to separate the breast as an organ from the chest wall. Manual segmentation or user-assisted interactive tools are inefficient, tedious, and error-prone, which is prohibitively impractical for processing large amounts of data from clinical trials. To address this challenge, we developed a fully automated and robust computerized segmentation method that intensively utilizes context information of breast MR imaging and the breast tissue's morphological characteristics to accurately delineate the breast and chest wall boundary. A critical component is the joint application of anisotropic diffusion and bilateral image filtering to enhance the edge that corresponds to the chest wall line (CWL) and to reduce the effect of adjacent non-CWL tissues. A CWL voting algorithm is proposed based on CWL candidates yielded from multiple sequential MRI slices, in which a CWL representative is generated and used through a dynamic time warping (DTW) algorithm to filter out inferior candidates, leaving the optimal one. Our method is validated by a representative dataset of 20 3D unilateral breast MRI scans that span the full range of the American College of Radiology (ACR) Breast Imaging Reporting and Data System (BI-RADS) fibroglandular density categorization. A promising performance (average overlay percentage of 89.33%) is observed when the automated segmentation is compared to manually segmented ground truth obtained by an experienced breast imaging radiologist. The automated method runs time-efficiently at ~3 minutes for each breast MR image set (28 slices).

Bilateral mammographic tissue density asymmetry could be an important factor in assessing risk of developing breast cancer and improving the detection of the suspicious lesions. This study aims to assess whether fusion of the bilateral mammographic density asymmetrical information into a computer-aided detection (CAD) scheme could improve CAD performance in detecting mass-like breast cancers. A testing dataset involving 1352 full-field digital mammograms (FFDM) acquired from 338 cases was used. In this dataset, half (169) cases are positive containing malignant masses and half are negative. Two computerized schemes were first independently applied to process FFDM images of each case. The first single-image based CAD scheme detected suspicious mass regions on each image. The second scheme detected and computed the bilateral mammographic tissue density asymmetry for each case. A fusion method was then applied to combine the output scores of the two schemes. The CAD performance levels using the original CAD-generated detection scores and the new fusion scores were evaluated and compared using a free-response receiver operating characteristic (FROC) type data analysis method. By fusion with the bilateral mammographic density asymmetrical scores, the case-based CAD sensitivity was increased from 79.2% to 84.6% at a false-positive rate of 0.3 per image. CAD also cued more "difficult" masses with lower CAD-generated detection scores while discarded some "easy" cases. The study indicated that fusion between the scores generated by a single-image based CAD scheme and the computed bilateral mammographic density asymmetry scores enabled to increase mass detection sensitivity in particular to detect more subtle masses.

We designed a Content-Based Image Retrieval (CBIR) Computer-Aided Diagnosis (CADx) system to assist radiologists in characterizing masses on ultrasound images. The CADx system retrieves masses that are similar to a query mass from a reference library based on computer-extracted features that describe texture, width-to-height ratio, and posterior shadowing of a mass. Retrieval is performed with k nearest neighbor (k-NN) method using Euclidean distance similarity measure and Rocchio relevance feedback algorithm (RRF). In this study, we evaluated the similarity between the query and the retrieved masses with relevance feedback using our interactive CBIR CADx system. The similarity assessment and feedback were provided by experienced radiologists' visual judgment. For training the RRF parameters, similarities of 1891 image pairs obtained from 62 masses were rated by 3 MQSA radiologists using a 9-point scale (9=most similar). A leave-one-out method was used in training. For each query mass, 5 most similar masses were retrieved from the reference library using radiologists' similarity ratings, which were then used by RRF to retrieve another 5 masses for the same query. The best RRF parameters were chosen based on three simulated observer experiments, each of which used one of the radiologists' ratings for retrieval and relevance feedback. For testing, 100 independent query masses on 100 images and 121 reference masses on 230 images were collected. Three radiologists rated the similarity between the query and the computer-retrieved masses. Average similarity ratings without and with RRF were 5.39 and 5.64 on the training set and 5.78 and 6.02 on the test set, respectively. The average Az values without and with RRF were 0.86±0.03 and 0.87±0.03 on the training set and 0.91±0.03 and 0.90±0.03 on the test set, respectively. This study demonstrated that RRF improved the similarity of the retrieved masses.

Computer-aided diagnosis (CAD) that uses content based image retrieval (CBIR) strategies has became an important research area. This paper presents a retrieval strategy that automatically recovers mammography masses from a virtual repository of mammographies. Unlike other approaches, we do not attempt to segment masses but instead we characterize the regions previously selected by an expert. These regions are firstly curvelet transformed and further characterized by approximating the marginal curvelet subband distribution with a generalized gaussian density (GGD). The content based retrieval strategy searches similar regions in a database using the Kullback-Leibler divergence as the similarity measure between distributions. The effectiveness of the proposed descriptor was assessed by comparing the automatically assigned label with a ground truth available in the DDSM database.¹ A total of 380 masses with different shapes, sizes and margins were used for evaluation, resulting in a mean average precision rate of 89.3% and recall rate of 75.2% for the retrieval task.

Patients with chronic lymphocytic leukemia (CLL) have an increased frequency of axillary lymphadenopathy. Pretreatment CT scans can be used to upstage patients at the time of presentation and post-treatment CT scans can reduce the number of complete responses. In the current clinical workflow, the detection and diagnosis of lymph nodes is usually performed manually by examining all slices of CT images, which can be time consuming and highly dependent on the observer's experience. A system for automatic lymph node detection and measurement is desired. We propose a computer aided detection (CAD) system for axillary lymph nodes on CT scans in CLL patients. The lung is first automatically segmented and the patient's body in lung region is extracted to set the search region for lymph nodes. Multi-scale Hessian based blob detection is then applied to detect potential lymph nodes within the search region. Next, the detected potential candidates are segmented by fast level set method. Finally, features are calculated from the segmented candidates and support vector machine (SVM) classification is utilized for false positive reduction. Two blobness features, Frangi's and Li's, are tested and their free-response receiver operating characteristic (FROC) curves are generated to assess system performance. We applied our detection system to 12 patients with 168 axillary lymph nodes measuring greater than 10 mm. All lymph nodes are manually labeled as ground truth. The system achieved sensitivities of 81% and 85% at 2 false positives per patient for Frangi's and Li's blobness, respectively.

In this paper, we present a computer-aided prognosis (CAP) scheme that utilizes quantitatively derived image information to predict patient recurrent-free survival for lung cancers. Our scheme involves analyzing CT histograms to evaluate the volumetric distribution of CT values within pulmonary nodules. A variational Bayesian mixture modeling framework translates the image-derived features into an image-based risk score for predicting the patient recurrence-free survival. Using our dataset of 454 patients with NSCLC, we demonstrate the potential usefulness of the CAP scheme which can provide a quantitative risk score that is strongly correlated with prognostic factors.

Precise size measurement of enlarged lymph nodes is a significant indicator for diagnosing malignancy, follow-up and therapy monitoring of cancer diseases. The presence of diverse sizes and shapes, inhomogeneous enhancement and the adjacency to neighboring structures with similar intensities, make the segmentation task challenging. We present a semi-automatic approach requiring minimal user interactions to fast and robustly segment the enlarged lymph nodes. First, a stroke approximating the largest diameter of a specific lymph node is drawn manually from which a volume of interest (VOI) is determined. Second, Based on the statistical analysis of the intensities on the dilated stroke area, a region growing procedure is utilized within the VOI to create an initial segmentation of the target lymph node. Third, a rotatable spiral-scanning technique is proposed to resample the 3D boundary surface of the lymph node to a 2D boundary contour in a transformed polar image. The boundary contour is found by seeking the optimal path in 2D polar image with dynamic programming algorithm and eventually transformed back to 3D. Ultimately, the boundary surface of the lymph node is determined using an interpolation scheme followed by post-processing steps. To test the robustness and efficiency of our method, a quantitative evaluation was conducted with a dataset of 315 lymph nodes acquired from 79 patients with lymphoma and melanoma. Compared to the reference segmentations, an average Dice coefficient of 0.88 with a standard deviation of 0.08, and an average absolute surface distance of 0.54mm with a standard deviation of 0.48mm, were achieved.

This paper presents a novel approach in computer aided skin lesion segmentation of dermoscopic images. We apply spatial and color features in order to model the lesion growth pattern. The decomposition is done by repeatedly clustering pixels into dark and light sub-clusters. A novel tree structure based representation of the lesion growth pattern is constructed by matching every pixel sub-cluster with a node in the tree structure. This model provides a powerful framework to extract features and to train models for lesion segmentation. The model employed allows features to be extracted at multiple layers of the tree structure, enabling a more descriptive feature set. Additionally, there is no need for preprocessing such as color calibration or artifact disocclusion. Preliminary features (mean over RGB color channels) are extracted for every pixel over four layers of the growth pattern model and are used in association with radial distance as a spatial feature to segment the lesion. The resulting per pixel feature vectors of length 13 are used in a supervised learning model for estimating parameters and segmenting the lesion. A dataset containing 116 challenging images from dermoscopic atlases is used to validate the method via a 10-fold cross validation procedure. Results of segmentation are compared with six other skin lesion segmentation methods. Our method outperforms ve other methods and performs competitively with another method. We achieve a per-pixel sensitivity/specicity of 0.890 and 0.901 respectively.

Quantification of overall tumor area on bone scans may be a potential biomarker for treatment response assessment and has, to date, not been investigated. Segmentation of bone metastases on bone scans is a fundamental step for this response marker. In this paper, we propose a fully automated computerized method for the segmentation of bone metastases on bone scans, taking into account characteristics of different anatomic regions. A scan is first segmented into anatomic regions via an atlas-based segmentation procedure, which involves non-rigidly registering a labeled atlas scan to the patient scan. Next, an intensity normalization method is applied to account for varying levels of radiotracer dosing levels and scan timing. Lastly, lesions are segmented via anatomic regionspecific intensity thresholding. Thresholds are chosen by receiver operating characteristic (ROC) curve analysis against manual contouring by board certified nuclear medicine physicians. A leave-one-out cross validation of our method on a set of 39 bone scans with metastases marked by 2 board-certified nuclear medicine physicians yielded a median sensitivity of 95.5%, and specificity of 93.9%. Our method was compared with a global intensity thresholding method. The results show a comparable sensitivity and significantly improved overall specificity, with a p-value of 0.0069.

MRI has shown to have great potential in prostate cancer localization and grading, but interpreting those exams requires expertise that is not widely available. Therefore, CAD applications are being developed to aid radiologists in detecting prostate cancer. Existing CAD applications focus on the prostate as a whole. However, in clinical practice transition zone cancer and peripheral zone cancer are considered to have different appearances. In this paper we present zone-specific CAD, in addition to an atlas based segmentation technique which includes zonal segmentation. Our CAD system consists of a detection and a classification stage. Prior to the detection stage the prostate is segmented into two zones. After segmentation features are extracted. Subsequently a likelihood map is generated on which local maxima detection is performed. For each local maximum a region is segmented. In the classification stage additional shape features are calculated, after which the regions are classified. Validation was performed on 288 data sets with MR-guided biopsy results as ground truth. Freeresponse Receiver Operating Characteristic (FROC) analysis was used for statistical evaluation. The difference between whole-prostate and zone-specific CAD was assessed using the difference between the FROCs. Our results show that evaluating the two zones separately results in an increase in performance compared to whole-prostate CAD. The FROC curves at .1, 1 and 3 false positives have a sensitivity of 0.0, 0.55 and 0.72 for whole-prostate and 0.08, 0.57 and 0.80 for zone-specific CAD. The FROC curve of the zone-specific CAD also showed significantly better performance overall (p < 0.05).

A major challenge in the current computer-aided detection (CADe) of hepatocellular carcinomas (HCCs) in contrastenhanced hepatic CT is to reduce the number of false-positive (FP) detections while maintaining a high sensitivity level. In this paper, we propose a feature selection method based on a sequential forward floating selection procedure coupled with a linear discriminant analysis classifier to improve the classification performance in computerized detection of HCCs in contrast-enhanced hepatic CT. The proposed method selected the most relevant features that would maximize the partial area under the receiver-operating-characteristic (ROC) curve (partial AUC) value, which would essentially lead to the maximum classification performance in the computer-aided detection scheme in a clinical setting. The partial AUC value is defined as the normalized AUC value in the high sensitivity region of the ROC curve, which is of clinical importance. In order to test the performance of the proposed method, we compared it against the popular stepwise feature selection method based on Wilks' lambda and a recently developed maximal AUC feature selection for an HCC database (23 HCCs and 1279 non-HCCs). We extracted 88 morphologic, gray-level-based, and texture features from the segmented lesion candidate regions in the hepatic CT images. The proposed method selected 9 features and achieved 100% sensitivity at 5.5 FPs per patient. Experiments showed a significant improvement in the performance of the classifier with the proposed feature selection method over that with the popular stepwise feature selection based on Wilks' lambda (17.3 FPs per patient) and the maximal AUC feature selection (10.0 FPs per patient) in terms of AUC values and FP rates.

This paper proposes a novel and fast approach for automatic estimation of the fetal weights from 3D ultrasound data. Conventional manual approaches are time-consuming and involve inconsistence by different sonographers because of the difficulty to trace limb boundaries in complicated Ultrasound limb volumes. It takes up to 10 minutes to manually trace the surface borders of a 20cm long limb. Using our automatic approach, the time is significantly reduced to 2.1 seconds for measuring the weights based on the entire limb. Experiments with the automatic approach also show comparable standard deviation and limits of agreement to the manual approaches.

We are developing a computerized system for bladder segmentation on CTU, as a critical component for computer aided diagnosis of bladder cancer. A challenge for bladder segmentation is the presence of regions without contrast (NC) and filled with IV contrast (C). We are developing a Conjoint Level set Analysis and Segmentation System (CLASS) specifically for this application. CLASS performs a series of image processing tasks: preprocessing, initial segmentation, and 3D and 2D level set segmentation and post-processing, designed according to the characteristics of the bladder in CTU. The NC and the C regions of the bladder were segmented separately in CLASS. The final contour is obtained in the post-processing stage by the union of the NC and C contours. Seventy bladders (31 containing lesions, 24 containing wall thickening, and 15 normal) were segmented. The performance of CLASS was assessed by rating the quality of the contours on a 5-point scale (1= "very poor", 3= "fair", 5 = "excellent"). For the 53 partially contrast-filled bladders, the average quality ratings for the 53 NC and 53 C regions were 4.0±0.7 and 4.0±1.0, respectively. 46 NC and 41 C regions were given quality ratings of 4 or above. Only 2 NC and 5 C regions had ratings under 3. The average quality ratings for the remaining 12 completely no contrast (NC) and 5 completely contrast-filled (C) bladder contours were 3.3±1.0 and 3.4±0.5, respectively. After combining the NC and C contours for each of the 70 bladders, 46 had quality ratings of 4 or above. Only 4 had ratings under 3. The average quality rating was 3.8±0.7. The results demonstrate the potential of CLASS for automated segmentation of the bladder.

Optical coherence tomography (OCT) is a light-based, high resolution imaging technique to guide stent deployment procedure for stenosis. OCT can accurately differentiate the most superficial layers of the vessel wall as well as stent struts and the vascular tissue surrounding them. In this paper, we automatically detect the struts of coronary stents present in OCT sequences. We propose a novel method to detect the strut shadow zone and accurately segment and reconstruct the strut in 3D. The estimation of the position of the strut shadow zone is the key requirement which enables the strut segmentation. After identification of the shadow zone we use probability map to estimate stent strut positions. This method can be applied to cross-sectional OCT images to detect the struts. Validation is performed using simulated data as well as in four in-vivo OCT sequences and the accuracy of strut detection is over 90%. The comparison against manual expert segmentation demonstrates that the proposed strut identification is robust and accurate.

Intravascular optical coherence tomography (IVOCT) provides very high resolution cross-sectional image sequences of vessels. It has been rapidly accepted for stent implantation and its follow up evaluation. Given the large amount of stent struts in a single image sequence, only automated detection methods are feasible. In this paper, we present an automated stent strut detection technique which requires neither lumen nor vessel wall segmentation. To detect strut-pixel candidates, both global intensity histograms and local intensity profiles of the raw polar images are used. Gaussian smoothing is applied followed by specified Prewitt compass filters to detect the trailing shadow of each strut. The shadow edge positions assist the strut-pixel candidates clustering. In the end, a 3D guide wire filter is applied to remove the guide wire from the detection results. For validation, two experts marked 6738 struts in 1021 frames in 10 IVOCT image sequences from a one-year follow up study. The struts were labeled as malapposed, apposed or covered together with the image quality (high, medium, low). The inter-observer agreement was 96%. The algorithm was validated for different combinations of strut status and image quality. Compared to the manual results, 93% of the struts were correctly detected by the new method. For each combination, the lowest accuracy was 88%, which shows the robustness towards different situations. The presented method can detect struts automatically regardless of the strut status or the image quality, which can be used for quantitative measurement, 3D reconstruction and visualization of the implanted stents.

Prenatal events such as intrauterine growth restriction have been shown to be associated with an increased thickness of abdominal aorta in the fetus. Therefore the measurement of abdominal aortic intima-media thickness (aIMT) has been recently considered a sensitive marker of artherosclerosis risk. To date measure of aortic diameter and of aIMT has been performed manually on US fetal images, thus being susceptible to intra- and inter- operator variability. This work introduces an automatic algorithm that identifies abdominal aorta and estimates its diameter and aIMT from videos recorded during routine third trimester ultrasonographic fetal biometry. Firstly, in each frame, the algorithm locates and segments the region corresponding to aorta by means of an active contour driven by two different external forces: a static vector field convolution force and a dynamic pressure force. Then, in each frame, the mean diameter of the vessel is computed, to reconstruct the cardiac cycle: in fact, we expect the diameter to have a sinusoidal trend, according to the heart rate. From the obtained sinusoid, we identify the frames corresponding to the end diastole and to the end systole. Finally, in these frames we assess the aIMT. According to its definition, we consider as aIMT the distance between the leading edge of the blood-intima interface, and the leading edge of the media-adventitia interface on the far wall of the vessel. The correlation between end-diastole and end-systole aIMT automatic and manual measures is 0.90 and 0.84 respectively.

Vessel segmentation is a fundamental step in an automated pulmonary embolism (PE) detection system. The purpose of this study is to improve the segmentation scheme for pulmonary vessels affected by PE and other lung diseases. We have developed a multiscale hierarchical vessel enhancement and segmentation (MHES) method for pulmonary vessel tree extraction based on the analysis of eigenvalues of Hessian matrices. However, it is difficult to segment the pulmonary vessels accurately under suboptimal conditions, such as vessels occluded by PEs, surrounded by lymphoid tissues or lung diseases, and crossing with other vessels. In this study, we developed a new vessel refinement method utilizing curved planar reformation (CPR) technique combined with optimal path finding method (MHES-CROP). The MHES segmented vessels straightened in the CPR volume was refined using adaptive gray level thresholding where the local threshold was obtained from least-square estimation of a spline curve fitted to the gray levels of the vessel along the straightened volume. An optimal path finding method based on Dijkstra's algorithm was finally used to trace the correct path for the vessel of interest. Two and eight CTPA scans were randomly selected as training and test data sets, respectively. Forty volumes of interest (VOIs) containing "representative" vessels were manually segmented by a radiologist experienced in CTPA interpretation and used as reference standard. The results show that, for the 32 test VOIs, the average percentage volume error relative to the reference standard was improved from 32.9±10.2% using the MHES method to 9.9±7.9% using the MHES-CROP method. The accuracy of vessel segmentation was improved significantly (p<0.05). The intraclass correlation coefficient (ICC) of the segmented vessel volume between the automated segmentation and the reference standard was improved from 0.919 to 0.988. Quantitative comparison of the MHES method and the MHES-CROP method with the reference standard was also evaluated by the Bland-Altman plot. This preliminary study indicates that the MHES-CROP method has the potential to improve PE detection.

Three-dimensional ultrasound (3DUS) provides non-invasive and precise measurements of carotid atherosclerosis that directly reflects arterial wall abnormalities that are thought to be related to stroke risk. Here we describe a threedimensional segmentation method based on the sparse field level set method to automate the segmentation of the mediaadventitia (MAB) and lumen-intima (LIB) boundaries of the common carotid artery from 3DUS images. To initiate the process, an expert chooses four anchor points on each boundary on a subset of transverse slices that are orthogonal to the axis of the artery. An initial surface is generated using the anchor points as initial guess for the segmentation. The MAB is segmented first using five energies: length minimization energy, local region-based energy, edge-based energy, anchor point-based energy, and local smoothness energy. Five energies are also used for the LIB segmentation: length minimization energy, local region-based energy, global region-based energy, anchor point-based energy, and boundary separation-based energy. The algorithm was evaluated with respect to manual segmentations on a slice-by-slice basis using 15 3DUS images. To generate results in this paper, inter-slice distance of 2 mm is used for the initialization. For the MAB and LIB segmentations, our method yielded Dice coefficients of more than 92% and sub-millimeter values for mean and maximum absolute distance errors. Our method also yielded a vessel wall volume error of 7.1% ± 3.4%. The realization of a semi-automated algorithm will aid in the translation of 3DUS measurements to clinical research for the rapid, non-invasive, and economical monitoring of atherosclerotic disease.

Chronic Obstructive Pulmonary Disease (COPD) is a chronic lung disease that is characterized by airflow limitation. COPD is clinically diagnosed and monitored using pulmonary function testing (PFT), which measures global inspiration and expiration capabilities of patients and is time-consuming and labor-intensive. It is becoming standard practice to obtain paired inspiration-expiration CT scans of COPD patients. Predicting the PFT results from the CT scans would alleviate the need for PFT testing. It is hypothesized that the change of the trachea during breathing might be an indicator of tracheomalacia in COPD patients and correlate with COPD severity. In this paper, we propose to automatically measure morphological changes in the trachea from paired inspiration and expiration CT scans and investigate the influence on COPD GOLD stage classification. The trachea is automatically segmented and the trachea shape is encoded using the lengths of rays cast from the center of gravity of the trachea. These features are used in a classifier, combined with emphysema scoring, to attempt to classify subjects into their COPD stage. A database of 187 subjects, well distributed over the COPD GOLD stages 0 through 4 was used for this study. The data was randomly divided into training and test set. Using the training scans, a nearest mean classifier was trained to classify the subjects into their correct GOLD stage using either emphysema score, tracheal shape features, or a combination. Combining the proposed trachea shape features with emphysema score, the classification performance into GOLD stages improved with 11% to 51%. In addition, an 80% accuracy was achieved in distinguishing healthy subjects from COPD patients.

Classification is widely used in the context of medical image analysis and in order to illustrate the mechanism of a classifier, we introduce the notion of an exaggerated image stereotype based on training data and trained classifier. The stereotype of some image class of interest should emphasize/exaggerate the characteristic patterns in an image class and visualize the information the employed classifier relies on. This is useful for gaining insight into the classification and serves for comparison with the biological models of disease. In this work, we build exaggerated image stereotypes by optimizing an objective function which consists of a discriminative term based on the classification accuracy, and a generative term based on the class distributions. A gradient descent method based on iterated conditional modes (ICM) is employed for optimization. We use this idea with Fisher's linear discriminant rule and assume a multivariate normal distribution for samples within a class. The proposed framework is applied to computed tomography (CT) images of lung tissue with emphysema. The synthesized stereotypes illustrate the exaggerated patterns of lung tissue with emphysema, which is underpinned by three different quantitative evaluation methods.

Tubes like Endotracheal (ET) tube used to maintain patient's airway and the Nasogastric (NG) tube used to feed the patient and drain contents of the stomach are very commonly used in Intensive Care Units (ICU). The placement of these tubes is critical for their proper functioning and improper tube placement can even be fatal. Bedside chest radiographs are considered the quickest and safest method to check the placement of these tubes. Tertiary ICU's typically generate over 250 chest radiographs per day to confirm tube placement. This paper develops a new fully automatic prototype computer-aided detection (CAD) system for tube detection on bedside chest radiographs. The core of the CAD system is the randomized algorithm which selects tubes based on their average repeatability from seed points. The CAD algorithm is designed as a 5 stage process: Preprocessing (removing borders, histogram equalization, anisotropic filtering), Anatomy Segmentation (to identify neck, esophagus, abdomen ROI's), Seed Generation, Region Growing and Tube Selection. The preliminary evaluation was carried out on 64 cases. The prototype CAD system was able to detect ET tubes with a True Positive Rate of 0.93 and False Positive Rate of 0.02/image and NG tubes with a True Positive Rate of 0.84 and False Positive Rate of 0.02/image respectively. The results from the prototype system show that it is feasible to automatically detect both tubes on chest radiographs, with the potential to significantly speed the delivery of imaging services while maintaining high accuracy.

In this study, we quantitatively characterize lung airway remodeling caused by smoking-related emphysema and Chronic Obstructive Pulmonary Disease (COPD), in low-dose CT scans. To that end, we established three groups of individuals: subjects with COPD (n=35), subjects with emphysema (n=38) and healthy smokers (n=28). All individuals underwent a low-dose CT scan, and the images were analyzed as described next. First the lung airways were segmented using a fast marching method and labeled according to its generation. Along each airway segment, cross-section images were resampled orthogonal to the airway axis. Next 128 rays were cast from the center of the airway lumen in each crosssection slice. Finally, we used an integral-based method, to measure lumen radius, wall thickness, mean wall percentage and mean peak wall attenuation on every cast ray. Our analysis shows that both the mean global wall thickness and the lumen radius of the airways of both COPD and emphysema groups were significantly different from those of the healthy group. In addition, the wall thickness change starts at the 3^rd airway generation in the COPD patients compared with emphysema patients, who display the first significant changes starting in the 2^nd generation. In conclusion, it is shown that airway remodeling happens in individuals suffering from either COPD or emphysema, with some local difference between both groups, and that we are able to detect and accurately quantify this process using images of low-dose CT scans.

Our purposes are to develop a conventional computer-aided diagnostic (CAD) scheme and a new fusion CAD scheme for the detection of lung nodules in multi-projection chest radiography, and to verify that information fused from the multi-projection chest radiography can greatly improve the performance of the conventional CAD scheme. The conventional CAD scheme processed each of the three projection images of a subject independently, and discarded the correlation information between the three images. The fusion CAD scheme registered all candidates detected by the conventional CAD scheme in the three images of a subject, and integrated the correlation information between the registered candidates to remove false positives. The CAD schemes were trained and evaluated on a database with 97 subjects. At the sensitivities of 70%, 65% and 60%, the conventional CAD scheme reported 20.4, 13.6 and 8.8 false positives per image, respectively, whereas the fusion CAD scheme reported 4.5, 2.8 and 1.2 false positives per image, respectively. The fusion of correlation information can markedly improve the performance of CAD scheme for lung nodule detection.

MRI perfusion images give information about regional lung function and can be used to detect pulmonary pathologies in cystic fibrosis (CF) children. However, manual assessment of the percentage of pathologic tissue in defined lung subvolumes features large inter- and intra-observer variation, making it difficult to determine disease progression consistently. We present an automated method to calculate a regional score for this purpose. First, lungs are located based on thresholding and morphological operations. Second, statistical shape models of left and right children's lungs are initialized at the determined locations and used to precisely segment morphological images. Segmentation results are transferred to perfusion maps and employed as masks to calculate perfusion statistics. An automated threshold to determine pathologic tissue is calculated and used to determine accurate regional scores. We evaluated the method on 10 MRI images and achieved an average surface distance of less than 1.5 mm compared to manual reference segmentations. Pathologic tissue was detected correctly in 9 cases. The approach seems suitable for detecting early signs of CF and monitoring response to therapy.

Computer-aided detection (CADe) has been investigated for assisting radiologists in detecting polyps in CT colonography (CTC). One of the major challenges in current CADe of polyps in CTC is to improve the specificity without sacrificing the sensitivity. We have developed several CADe schemes based on a massive-training framework with different nonlinear regression models such as neural network regression, support vector regression, and Gaussian process regression. Individual CADe schemes based on different nonlinear regression models, however, achieved comparable results. In this paper, we propose to use the AdaBoost algorithm to combine different regression models in CADe schemes for improving the specificity without sacrificing the sensitivity. To test the performance of the proposed approach, we compared it with individual regression models in the distinction between polyps and various types of false positives (FPs). Our CTC database consisted of 246 CTC datasets obtained from 123 patients in the supine and prone positions. The testing set contained 93 patients including 19 polyps in seven patients and 86 negative patients with 474 FPs produced by an original CADe scheme. The AdaBoost algorithm combining multiple massive-training regression models achieved a performance that was higher than each individual regression model, yielding a 94.7% (18/19) bypolyp sensitivity at an FP rate of 2.0 (188/93) per patient in a leave-one-lesion-out cross validation test.

Colon cancer is the third most commonly diagnosed type of cancer in the US. In recent years, however, early diagnosis and treatment have caused a significant rise in the five year survival rate. Preventive screening is often performed by colonoscopy (endoscopic inspection of the colon mucosa). Narrow Band Imaging (NBI) is a novel diagnostic approach highlighting blood vessel structures on polyps which are an indicator for future cancer risk. In this paper, we review our automated inter- and intra-observer independent system for the automated classification of polyps into hyperplasias and adenomas based on vessel structures to further improve the classification performance. To surpass the performance limitations we derive a novel vessel segmentation approach, extract 22 features to describe complex vessel topologies, and apply three feature selection strategies. Tests are conducted on 286 NBI images with diagnostically important and challenging polyps (10mm or smaller) taken from our representative polyp database. Evaluations are based on ground truth data determined by histopathological analysis. Feature selection by Simulated Annealing yields the best result with a prediction accuracy of 96.2% (sensitivity: 97.6%, specificity: 94.2%) using eight features. Future development aims at implementing a demonstrator platform to begin clinical trials at University Hospital Aachen.

Human colon has complex structures mostly because of the haustral folds. Haustral folds are thin flat protrusions on the colon wall, which inherently attached on the colon wall. These structures may complicate the shape analysis for computer-aided detection of colonic polyps (CADpolyp); however, they can serve as solid reference during image interpretation in computed tomographic colonography (CTC). Therefore, in this study, based on a clear model of the haustral fold boundaries, we employ level set method to automatically segment the fold surfaces. We believe the segmented folds have the potential to significantly benefit various post-procedures in CTC, e.g., supine-prone registration, synchronized image interpretation, automatic polyp matching, CADpolyp, teniae coli extraction, etc. For the first time, with assistance from physician experts, we established the ground truth of haustral fold boundaries of 15 real patient data from two medical centers, based on which we evaluated our algorithm. The results demonstrated that about 92.7% of the folds are successfully detected. Furthermore, we explored the segmented area ratio (SAR), i.e., the ratio between the areas of the intersection and the union of the expert-drawn and the automatically-segmented folds, to measure the accuracy of the segmentation algorithm. The averaged result of SAR=86.2% shows a good match between the ground truth and our segmentation results.

Conventional single-energy computed tomography colonography (CTC) tends to miss polyps 6 - 9 mm in size and flat lesions. Dual-energy CTC (DE-CTC) provides more complete information about the chemical composition of tissue than does conventional CTC. We developed an automated computer-aided detection (CAD) scheme for detecting colorectal lesions in which dual-energy features were used to identify different bowel materials and their partial-volume artifacts. Based on these features, the dual-energy CAD (DE-CAD) scheme extracted the region of colon by use of a lumen-tracking method, detected lesions by use of volumetric shape features, and reduced false positives by use of a statistical classifier. For validation, 20 patients were prepared for DE-CTC by use of reduced bowel cleansing and orally administered fecal tagging with iodine and/or barium. The DE-CTC was performed in dual positions by use of a dual-energy CT scanner (SOMATOM Definition, Siemens) at 140 kVp and 80 kVp energy levels. The lesions identified by subsequent same-day colonoscopy were correlated with the DE-CTC data. The detection accuracies of the DE-CAD and conventional CAD schemes were compared by use of leave-one-patient-out evaluation and a bootstrap analysis. There were 25 colonoscopy-confirmed lesions: 22 were 6 - 9 mm and 3 were flat lesions ≥10 mm in size. The DE-CAD scheme detected the large flat lesions and 95% of the 6 - 9 mm lesions with 9.9 false positives per patient. The improvement in detection accuracy by the DE-CAD was statistically significant.

Computed tomographic colonography (CTC) is a minimally invasive technique for colonic polyps and cancer screening. The marginal artery of the colon, also known as the marginal artery of Drummond, is the blood vessel that connects the inferior mesenteric artery with the superior mesenteric artery. The marginal artery runs parallel to the colon for its entire length, providing the blood supply to the colon. Detecting the marginal artery may benefit computer-aided detection (CAD) of colonic polyp. It can be used to identify teniae coli based on their anatomic spatial relationship. It can also serve as an alternative marker for colon localization, in case of colon collapse and inability to directly compute the endoluminal centerline. This paper proposes an automatic method for marginal artery detection on CTC. To the best of our knowledge, this is the first work presented for this purpose. Our method includes two stages. The first stage extracts the blood vessels in the abdominal region. The eigenvalue of Hessian matrix is used to detect line-like structures in the images. The second stage is to reduce the false positives in the first step. We used two different masks to exclude the false positive vessel regions. One is a dilated colon mask which is obtained by colon segmentation. The other is an eroded visceral fat mask which is obtained by fat segmentation in the abdominal region. We tested our method on a CTC dataset with 6 cases. Using ratio-of-overlap with manual labeling of the marginal artery as the standard-of-reference, our method yielded true positive, false positive and false negative fractions of 89%, 33%, 11%, respectively.

Accurate and objective evaluation of vertebral body deformations represents an important part of the clinical diagnostics and therapy of pathological conditions affecting the spine. Although modern clinical practice is oriented towards threedimensional (3D) imaging techniques, the established methods for the evaluation of vertebral body deformations are based on measurements in two-dimensional (2D) X-ray images. In this paper, we propose a method for automatic measurement of vertebral body deformations in computed tomography (CT) images that is based on efficient modeling of the vertebral body shape with a 3D parametric model. By fitting the 3D model to the vertebral body in the image, quantitative description of normal and pathological vertebral bodies is obtained from the value of 25 parameters of the model. The evaluation of vertebral body deformations is based on the distance of the observed vertebral body from the distribution of the parameter values of normal vertebral bodies in the parametric space. The distribution is obtained from 80 normal vertebral bodies in the training data set and verified with eight normal vertebral bodies in the control data set. The statistically meaningful distance of eight pathological vertebral bodies in the study data set from the distribution of normal vertebral bodies in the parametric space shows that the parameters can be used to successfully model vertebral body deformations in 3D. The proposed method may therefore be used to assess vertebral body deformations in 3D or provide clinically meaningful observations that are not available when using 2D methods that are established in clinical practice.

In an attempt to improve the visualisation techniques for diagnosis and treatment of musculoskeletal injuries, we present a novel image fusion method for a pixel-wise fusion of CT and MR images. We focus on the spine and it's related diseases including osteophyte growth, degenerate disc disease and spinal stenosis. This will have benefit to the 50-75% of people who suffer from back pain, which is the reason for 1.8% of all hospital stays in the United States.¹ Pre-registered CT and MR image pairs were used. Rigid registration was performed based on soft tissue correspondence. A pixel-wise image fusion algorithm has been designed to combine CT and MR images into a single image. This is accomplished by minimizing an energy functional using a Graph Cut approach. The functional is formulated to balance the similarity between the resultant image and the CT image as well as between the resultant image and the MR image. Furthermore the variational smoothness of the resultant image is considered in the energy functional (to enforce natural transitions between pixels). The results have been validated based on the amount of significant detail preserved in the final fused image. Based on bone cortex and disc / spinal cord areas, 95% of the relevant MR detail and 85% of the relevant CT detail was preserved. This work has the potential to aid in patient diagnosis, surgery planning and execution along with post operative follow up.

The early detection of bone metastases is important for determining the prognosis and treatment of a patient. We developed a CAD system which detects sclerotic bone metastases in the spine on CT images. After the spine is segmented from the image, a watershed algorithm detects lesion candidates. The over-segmentation problem of the watershed algorithm is addressed by the novel incorporation of a graph-cuts driven merger. 30 quantitative features for each detection are computed to train a support vector machine (SVM) classifier. The classifier was trained on 12 clinical cases and tested on 10 independent clinical cases. Ground truth lesions were manually segmented by an expert. The system prior to classification detected 87% (72/83) of the manually segmented lesions with volume greater than 300 mm³. On the independent test set, the sensitivity was 71.2% (95% confidence interval (63.1%, 77.3%)) with 8.8 false positives per case.

Spinal bone lesion detection is a challenging and important task in cancer diagnosis and treatment monitoring. In this paper we present a method for fully-automatic osteolytic spinal bone lesion detection from 3D CT data. It is a multi-stage approach subsequently applying multiple discriminative models, i.e., multiple random forests, for lesion candidate detection and rejection to an input volume. For each detection stage an internal control mechanism ensures maintaining sensitivity on unseen true positive lesion candidates during training. This way a pre-defined target sensitivity score of the overall system can be taken into account at the time of model generation. For a lesion not only the center is detected but also, during post-processing, its spatial extension along the three spatial axes defined by the surrounding vertebral body's local coordinate system. Our method achieves a cross-validated sensitivity score of 75% and a mean false positive rate of 3.0 per volume on a data collection consisting of 34 patients with 105 osteolytic spinal bone lesions. The median sensitivity score is 86% at 2.0 false positives per volume.

Adolescent idiopathic scoliosis (AIS) is a deformity of the spine manifested by asymmetry and deformities of the external surface of the trunk. Classification of scoliosis deformities according to curve type is used to plan management of scoliosis patients. Currently, scoliosis curve type is determined based on X-ray exam. However, cumulative exposure to X-rays radiation significantly increases the risk for certain cancer. In this paper, we propose a robust system that can classify the scoliosis curve type from non invasive acquisition of 3D trunk surface of the patients. The 3D image of the trunk is divided into patches and local geometric descriptors characterizing the surface of the back are computed from each patch and forming the features. We perform the reduction of the dimensionality by using Principal Component Analysis and 53 components were retained. In this work a multi-class classifier is built with Least-squares support vector machine (LS-SVM) which is a kernel classifier. For this study, a new kernel was designed in order to achieve a robust classifier in comparison with polynomial and Gaussian kernel. The proposed system was validated using data of 103 patients with different scoliosis curve types diagnosed and classified by an orthopedic surgeon from the X-ray images. The average rate of successful classification was 93.3% with a better rate of prediction for the major thoracic and lumbar/thoracolumbar types.

Detection of malignancy from histopathological images of breast cancer is a labor-intensive and error-prone process. To streamline this process, we present an efficient Computer Aided Diagnostic system that can differentiate between cancerous and non-cancerous H&E (hemotoxylin&eosin) biopsy samples. Our system uses novel textural, topological and morphometric features taking advantage of the special patterns of the nuclei cells in breast cancer histopathological images. We use a Support Vector Machine classifier on these features to diagnose malignancy. In conjunction with the maximum relevance - minimum redundancy feature selection technique, we obtain high sensitivity and specificity. We have also investigated the effect of image compression on classification performance.

We present a method for the computer-aided histopathological grading of follicular lymphoma (FL) images based on a multi-scale feature analysis. We analyze FL images using cell-graphs to characterize the structural organization of the cells in tissues. Cell-graphs represent histopathological images with undirected and unweighted graphs wherein the cytological components constitute the graph nodes and the approximate adjacencies of the components are represented with edges. Using the features extracted from nuclei- and cytoplasm-based cell-graphs, a classifier defines the grading of the follicular lymphoma images. The performance of this system is comparable to that of our recently developed system that characterizes higher-level semantic description of tissues using model-based intermediate representation (MBIR) and color-textural analysis. When tested with three different classifiers, the combination of cell-graph based features with the MBIR and color-textural features followed by a multi-scale feature selection is shown to achieve considerably higher classification accuracies than any set of these feature sets can achieve separately.

DNA Image Cytometry is a method for non-invasive cancer diagnosis which measures the DNA content of Feulgen-stained nuclei. DNA content is measured using a microscope system equipped with a digital camera as a densitometer and estimating the DNA content from the absorption of light when passing through the nuclei. However, a DNA Image Cytometry measurement is only valid if each nucleus is only measured once. To assist the user in preventing multiple measurements of the same nucleus, we have developed a unique digital identifier for the characterization of Feulgen-stained nuclei, the so called Nucleus Fingerprint. Only nuclei with a new fingerprint can be added to the measurement. This fingerprint is based on basic nucleus features, the contour of the nucleus and the spatial relationship to nuclei in the vicinity. Based on this characterization, a classifier for testing two nuclei for identity is presented. In a pairwise comparison of ≈40000 pairs of mutually different nuclei, 99.5% were classified as different. In another 450 tests, the fingerprints of the same nucleus recorded a second time were in all cases judged identical. We therefore conclude that our Nucleus Fingerprint approach robustly prevents the repeated measurement of nuclei in DNA Image Cytometry.

Periapical lesion is a common disease in oral health. While many studies have been devoted to image-based diagnosis of periapical lesion, these studies usually require clinicians to perform the task. In this paper we investigate the automatic solutions toward periapical lesion classification using quantized texture analysis. Specifically, we adapt the bag-of-visual-words model for periapical root image representation, which captures the texture information by collecting local patch statistics. Then we investigate several similarity measure approaches with the K-nearest neighbor (KNN) classifier for the diagnosis task. To evaluate these classifiers we have collected a digitized oral X-Ray image dataset from 21 patients, resulting 139 root images in total. The extensive experimental results demonstrate that the KNN classifier based on the bagof- words model can achieve very promising performance for periapical lesion classification.

The ratio between the amount of adipose and skeletal muscle tissue is an important determinant of metabolic health. Recent developments in MRI technology allow whole body scans to be performed for accurate assessment of body composition. In the present study, a total of 194 participants underwent a 2-point Dixon MRI sequence of the whole body. A fully automated image segmentation method quantifies the amount of adipose and skeletal muscle tissue by applying standard image processing techniques including thresholding, region growing and morphological operators. The adipose tissue is further divided into subcutaneous and visceral adipose tissue by using statistical shape models. All images were visually inspected. The quantitative analysis was performed on 44 whole-body MRI data using manual segmentations as ground truth data. We achieved 3.3% and 6.3% of relative volume difference between the manual and automated segmentation of subcutaneous and visceral adipose tissue, respectively. The validation of skeletal muscle tissue segmentation resulted in a relative volume difference of 7.8 ± 4.2% and a volumetric overlap error of 6.4 ± 2.3 %. To our knowledge, we are first to present a fully automated method which quantifies adipose and skeletal muscle tissue in whole-body MRI data. Due to the fully automated approach, results are deterministic and free of user bias. Hence, the software can be used in large epidemiological studies for assessing body fat distribution and the ratio of adipose to skeletal muscle tissue in relation to metabolic disease risk.

Magnetically-guided capsule endoscopy (MGCE) is a nascent technology with the goal to allow the steering of a capsule endoscope inside a water filled stomach through an external magnetic field. We developed a classification cascade for MGCE images with groups images in semantic and topological categories. Results can be used in a post-procedure review or as a starting point for algorithms classifying pathologies. The first semantic classification step discards over-/under-exposed images as well as images with a large amount of debris. The second topological classification step groups images with respect to their position in the upper gastrointestinal tract (mouth, esophagus, stomach, duodenum). In the third stage two parallel classifications steps distinguish topologically different regions inside the stomach (cardia, fundus, pylorus, antrum, peristaltic view). For image classification, global image features and local texture features were applied and their performance was evaluated. We show that the third classification step can be improved by a bubble and debris segmentation because it limits feature extraction to discriminative areas only. We also investigated the impact of segmenting intestinal folds on the identification of different semantic camera positions. The results of classifications with a support-vector-machine show the significance of color histogram features for the classification of corrupted images (97%). Features extracted from intestinal fold segmentation lead only to a minor improvement (3%) in discriminating different camera positions.

Retinal vasculature is one of the most important anatomical structures in digital retinal photographs. Accurate segmentation of retinal blood vessels is an essential task in automated analysis of retinopathy. This paper presents a new and effective vessel segmentation algorithm that features computational simplicity and fast implementation. This method uses morphological pre-processing to decrease the disturbance of bright structures and lesions before vessel extraction. Next, a vessel probability map is generated by computing the eigenvalues of the second derivatives of Gaussian filtered image at multiple scales. Then, the second order local entropy thresholding is applied to segment the vessel map. Lastly, a rule-based decision step, which measures the geometric shape difference between vessels and lesions is applied to reduce false positives. The algorithm is evaluated on the low-resolution DRIVE and STARE databases and the publicly available high-resolution image database from Friedrich-Alexander University Erlangen-Nuremberg (Germany). The proposed method achieved comparable performance to state of the art unsupervised vessel segmentation methods with a competitive faster speed on the DRIVE and STARE databases. For the high resolution fundus image database, the proposed algorithm outperforms an existing approach both on performance and speed. The efficiency and robustness make the blood vessel segmentation method described here suitable for broad application in automated analysis of retinal images.

Retinal blood vessels show morphologic modifications in response to various retinopathies. However, the specific responses exhibited by arteries and veins may provide a precise diagnostic information, i.e., a diabetic retinopathy may be detected more accurately with the venous dilatation instead of average vessel dilatation. In order to analyze the vessel type specific morphologic modifications, the classification of a vessel network into arteries and veins is required. We previously described a method for identification and separation of retinal vessel trees; i.e. structural mapping. Therefore, we propose the artery-venous classification based on structural mapping and identification of color properties prominent to the vessel types. The mean and standard deviation of each of green channel intensity and hue channel intensity are analyzed in a region of interest around each centerline pixel of a vessel. Using the vector of color properties extracted from each centerline pixel, it is classified into one of the two clusters (artery and vein), obtained by the fuzzy-C-means clustering. According to the proportion of clustered centerline pixels in a particular vessel, and utilizing the artery-venous crossing property of retinal vessels, each vessel is assigned a label of an artery or a vein. The classification results are compared with the manually annotated ground truth (gold standard). We applied the proposed method to a dataset of 15 retinal color fundus images resulting in an accuracy of 88.28% correctly classified vessel pixels. The automated classification results match well with the gold standard suggesting its potential in artery-venous classification and the respective morphology analysis.

Radiofrequency ablation is a promising procedure for treating atrial fibrillation (AF) that relies on accurate lesion delivery in the left atrial (LA) wall for success. Late Gadolinium Enhancement MRI (LGE MRI) at three months post-ablation has proven effective for noninvasive assessment of the location and extent of scar formation, which are important factors for predicting patient outcome and planning of redo ablation procedures. We have developed an algorithm for automatic classification in LGE MRI of scar tissue in the LA wall and have evaluated accuracy and consistency compared to manual scar classifications by expert observers. Our approach clusters voxels based on normalized intensity and was chosen through a systematic comparison of the performance of multivariate clustering on many combinations of image texture. Algorithm performance was determined by overlap with ground truth, using multiple overlap measures, and the accuracy of the estimation of the total amount of scar in the LA. Ground truth was determined using the STAPLE algorithm, which produces a probabilistic estimate of the true scar classification from multiple expert manual segmentations. Evaluation of the ground truth data set was based on both inter- and intra-observer agreement, with variation among expert classifiers indicating the difficulty of scar classification for a given a dataset. Our proposed automatic scar classification algorithm performs well for both scar localization and estimation of scar volume: for ground truth datasets considered easy, variability from the ground truth was low; for those considered difficult, variability from ground truth was on par with the variability across experts.

We propose a robust and fully automatic algorithm which computes the 2D echocardiography measurements recommended by America Society of Echocardiography. The algorithm employs knowledge-based imaging technologies which can learn the expert's knowledge from the training images and expert's annotation. Based on the models constructed from the learning stage, the algorithm searches initial location of the landmark points for the measurements by utilizing heart structure of left ventricle including mitral valve aortic valve. It employs the pseudo anatomic M-mode image generated by accumulating the line images in 2D parasternal long axis view along the time to refine the measurement landmark points. The experiment results with large volume of data show that the algorithm runs fast and is robust comparable to expert.

A significant cause of coronary artery disease is the coronary atherosclerosis which leads to stenosis of coronary arteries. It has been shown in recent studies, using intravascular ultrasound and contrast-enhanced CT, that early atherosclerosis causes positive coronary artery remodeling, defined as increases in the cross-sectional area. It is hypothesized that detection of artery remodeling using non-contrast CT can be an important factor in sub-clinical assessment of cardiac risk for asymptomatic subjects. However, measuring remodeling in coronary arteries in non-contrast CT images is a challenging task because coronary arteries are small and the intensity of coronary arteries is similar to that of surrounding tissues. Automatic segmentation algorithms that have been successful in segmenting coronary arteries in contrast-enhanced images do not perform well. To overcome these difficulties, we developed an interactive application to enable effective measurement of coronary artery remodeling in non-contrast CT images. This application is an extension to the 3D Slicer image analysis platform. It allows users to visualize and trace the centerline of arteries in cross sectional views. The artery centerlines are displayed in a three dimensional view overlaid on the original image volume and color-coded according to the artery labels. Using this 3D artery model, the user can sample the cross-sectional area of the arteries at selected points for remodeling assessment. Initial validation has demonstrated the effectiveness of this method. A pilot study also showed positive correlation of large coronary artery remodeling with highest lifetime risks. Further evaluation is underway using larger study size and more measurement points.

In this paper, a computer aided diagnosis (CAD) system for automatic detection of focal cortical dysplasia (FCD) on T1-weighted MRI is proposed. We introduce a new set of differential cluster-wise features comparing local differences of the candidate lesional area with its surroundings and other GM/WM boundaries. The local differences are measured in a distributional sense using χ² distances. Finally, a Support Vector Machine (SVM) classifier is used to classify the clusters. Experimental results show an 88% lesion detection rate with only 1.67 false positive clusters per subject. Also, the results show that using additional differential features clearly outperforms the result using only absolute features.

In imaging studies of neonates, particularly in the clinical setting, diusion tensor imaging-based tractography is typically unreliable due to the use of fast acquisition protocols that yield low resolution and signal-to-noise ratio (SNR). These image acquisition protocols are implemented with the aim of reducing displacement artifacts that may be produced by the movement of the neonate's head during the scanning session. In addition, axons are not yet fully myelinated in these subjects. As a result, the water molecules' movements are not as constrained as in older brains, making it even more dicult to dene structure by means of diusion proles. Here, we introduce a post-processing method that overcomes some of the diculties described above, allowing the determination of reliable tracts in newborns. We test our method using neonatal data and in particular, we successfully extract some of the limbic, association and commissural bers, all of which are typically dicult to obtain by direct tractography. The method is further validated through visual inspection by expert pediatric neuroradiologists.

We propose a system to automatically detect cerebral aneurysms in 3D X-ray rotational angiography images (3D-RA), magnetic resonance angiography images (MRA) and computed tomography angiography images (CTA). After image normalization, initial candidates are found by applying a blob-enhancing filter on the data sets. Clusters are computed by a modified k-means algorithm. A post-processing step reduces the false positive (FP) rate on the basis of computed features. This is implemented as a rule-based system that is adapted according to the modality. In MRA, clusters are excluded that are not neighbored to a vessel. As a final step, FP are further reduced by applying a threshold classification on a feature. Our method was tested on 93 angiographic data sets containing aneurysm and non-aneurysm cases. We achieved 95 % sensitivity with an average rate of 2.6 FP per data set (FP/DS) in case of 3D-RA, 89 % sensitivity at 6.6 FP/DS for MRA and 95 % sensitivity at 37.6 FP/DS with CTA, respectively. We showed that our post-processing approach eliminates FP in MRA with only a slight decrease of sensitivity. In contrast to other approaches, our algorithm does not require a vessel segmentation and does not require training of distributional properties.

Gleason patterns of prostate cancer histopathology, characterized primarily by morphological and architectural attributes of histological structures (glands and nuclei), have been found to be highly correlated with disease aggressiveness and patient outcome. Gleason patterns 4 and 5 are highly correlated with more aggressive disease and poorer patient outcome, while Gleason patterns 1-3 tend to reflect more favorable patient outcome. Because Gleason grading is done manually by a pathologist visually examining glass (or digital) slides, subtle morphologic and architectural differences of histological attributes may result in grading errors and hence cause high inter-observer variability. Recently some researchers have proposed computerized decision support systems to automatically grade Gleason patterns by using features pertaining to nuclear architecture, gland morphology, as well as tissue texture. Automated characterization of gland morphology has been shown to distinguish between intermediate Gleason patterns 3 and 4 with high accuracy. Manifold learning (ML) schemes attempt to generate a low dimensional manifold representation of a higher dimensional feature space while simultaneously preserving nonlinear relationships between object instances. Classification can then be performed in the low dimensional space with high accuracy. However ML is sensitive to the samples contained in the dataset; changes in the dataset may alter the manifold structure. In this paper we present a manifold regularization technique to constrain the low dimensional manifold to a specific range of possible manifold shapes, the range being determined via a statistical shape model of manifolds (SSMM). In this work we demonstrate applications of the SSMM in (1) identifying samples on the manifold which contain noise, defined as those samples which deviate from the SSMM, and (2) accurate out-of-sample extrapolation (OSE) of newly acquired samples onto a manifold constrained by the SSMM. We demonstrate these applications of the SSMM in the context of distinguishing between Gleason patterns 3 and 4 using glandular morphologic features in a prostate histopathology dataset of 58 patient studies. Identifying and eliminating noisy samples from the manifold via the SSMM results in a statistically significant improvement in classification accuracy (CA), 93.0 ± 1.0% with removal of noisy samples compared to a CA of 90.9 ± 1.1% without removal of samples. The use of the SSMM for OSE of new independent test instances also shows statistically significant improvement in CA, 87.1±0.8% with the SSMM compared to 85.6±0.1% without the SSMM. Similar improvements were observed for the synthetic Swiss Roll and Helix datasets.

There is a need for identifying quantitative imaging (e.g. MRI) signatures for prostate cancer (CaP), so that computer-aided diagnostic methods can be trained to detect disease extent in vivo. Determining CaP extent on in vivo MRI is difficult to do; however, with the availability of ex vivo surgical whole mount histological sections (WMHS) for CaP patients undergoing radical prostatectomy, co-registration methods can be applied to align and map disease extent onto pre-operative MR imaging from the post-operative histology. Yet obtaining digitized images of WHMS for co-registration with the pre-operative MRI is cumbersome since (a) most digital slide scanners are unable to accommodate the entire section, and (b) significant technical expertise is required for whole mount slide preparation. Consequently, most centers opt to construct quartered sections of each histology slice. Prior to co-registration with MRI, however, these quartered sections need to be digitally stitched together to reconstitute a digital, pseudo WMHS. Histostitcher© is an interactive software program that uses semi-automatic registration tools to digitally stitch quartered sections into pseudo WMHS. Histostitcher© was originally developed using the GUI tools provided by the Matlab programming interface, but the clinical use was limited due to the inefficiency of the interface. The limitations of the Matlab based GUI include (a) an inability to edit the fiducials, (b) the rendering being extremely slow, and (c) lack of interactive and rapid visualization tools. In this work, Histostitcher© has been integrated into the eXtensible Imaging Platform (XIP ^TM ) framework (a set of libraries containing functionalities for analyzing and visualizing medical image data). XIP ^TM lends the stitching tool much greater flexibility and functionality by (a) allowing interactive and seamless navigation through the full resolution histology images, (b) the ability to easily add, edit, or remove fiducials and annotations in order to register the quadrants and map the disease extent. In this work, we showcase examples of digital stitching of quartered histological sections into pseudo-WHMS using Histostitcher © via the new XIP ^TM interface. This tool will be particularly useful in clinical trials and large cohort studies where a quick, interactive way of digitally reconstructing pseudo WMHS is required.

Orally administered tagging agents are usually used in CT colonography (CTC) to differentiate residual bowel content from native colonic structure. However, the high-density contrast agents tend to introduce the scatter effect on neighboring soft tissues and elevate their observed CT attenuation values toward that of the tagged materials (TMs), which may result in an excessive electronic colon cleansing (ECC) where pseudo-enhanced soft tissues are incorrectly identified as TMs. To address this issue, we integrated a scale-based scatter correction as a preprocessing procedure into our previous ECC pipeline based on the maximum a posteriori expectation-maximization (MAP-EM) partial volume segmentation. The newly proposed ECC scheme takes into account both scatter effect and partial volume effect that commonly appear in CTC images. We evaluated the new method with 10 patient CTC studies and found improved performance. Our results suggest that the proposed strategy is effective with potentially significant benefits for both clinical CTC examinations and automatic computer-aided detection (CAD) of colon polyps.

In this paper, we propose an automated incision line determination method for virtual unfolded view generation of the stomach from 3D abdominal CT images. The previous virtual unfolding methods of the stomach required a lot of manual operations such as determination of the incision line, which heavily tasks an operator. In general, an incision line along the greater curvature of the stomach is used for making pathological specimen. In our method, an incision line is automatically determined by projecting a centerline of the stomach onto the gastric surface from a projection line. The projection line is determined by using positions of the cardia and the pylorus, that can be easily specified by two mouse clicks. The process of our method is performed as follows. We extract the stomach region using a thresholding and a labeling processes. We apply a thinning process to the stomach region, and then we extract the longest line from the result of the thinning process. We obtain a centerline of the stomach region by smoothing the longest line by using a Bezier curve. The incision line is calculated by projecting the centerline onto the gastric surface from the projection line. We applied the proposed method to 19 cases of CT images. We automatically determined incision lines. Experimintal results showed our method was able to determine incision lines along the greater curvature for most of 19 cases.

Phantom experiments are useful and frequently used in validating algorithms or techniques in applications where it is difficult or impossible to generate accurate ground-truth. In this work we present a phantom design and experiments to validate our colonoscopy tracking algorithms, that serve to keep both virtual colonoscopy and optical colonoscopy images aligned (in location and orientation). We describe the construction of two phantoms, capable of respectively moving along a straight and a curved path. The phantoms are motorized so as to be able to move at a near constant speed. Experiments were performed at three speeds: 10, 15 and 20mm/sec, to simulate motion velocities during colonoscopy procedures. The average velocity error was within 3mm/sec in both straight and curved phantoms. Displacement error was within 7mm over a total distance of 288mm in the straight phantom, and less than 7mm over 287mm in the curved phantom. Multiple trials were performed of each experiment(and their errors averaged) to ensure repeatability.

Bladder cancer is one of the most common cancers in the western world. The diagnosis in Germany is based on the visual inspection of the bladder. This inspection performed with a cystoscope is a challenging task as some kinds of abnormal tissues do not differ much in their appearance from their surrounding healthy tissue. Fluorescence Cystoscopy has the potential to increase the detection rate. A liquid marker introduced into the bladder in advance of the inspection is concentrated in areas with high metabolism. Thus these areas appear as bright "glowing". Unfortunately, the fluorescence image contains besides the glowing of the suspicious lesions no more further visual information like for example the appearance of the blood vessels. A visual judgment of the lesion as well as a precise treatment has to be done using white light illumination. Thereby, the spatial information of the lesion provided by the fluorescence image has to be guessed by the clinical expert. This leads to a time consuming procedure due to many switches between the modalities and increases the risk of mistreatment. We introduce an automatic approach, which detects and segments any suspicious lesion in the fluorescence image automatically once the image was classified as a fluorescence image. The area of the contour of the detected lesion is transferred to the corresponding white light image and provide the clinical expert the spatial information of the lesion. The advantage of this approach is, that the clinical expert gets the spatial and the visual information of the lesion together in one image. This can save time and decrease the risk of an incomplete removal of a malign lesion.

Automatic detection of hepatocellular carcinoma (HCC) from 3D CT images effectively reduces interpretation work. Several detection methods have been proposed. However, there still remains a tough problem of adaptation detection methods to a wide range of tumor sizes, especially to small nodules, since it is difficult to distinguish tumors from other structures, including noise. Although the level set method (LS) is a powerful tool for detecting objects with arbitrary topology, it is still poor at detecting small nodules due to low contrast. To detect small nodules, early phase images are useful since low contrast in the late phase causes miss-detection of some small nodules. Nevertheless, conventional methods using early phase images face two problems: one is failure to extract small nodules due to low contrast even in early phase images, and the other is false-positive (FP) detection of vessels adjacent to tumors. In this paper, a new robust detection method adapted to the wide range of tumor sizes has been proposed that uses only early phase images. To overcome these two problems, our method consists of two techniques. One is regularizing surface evolution used in LS by applying a new HCC filter that can enhance both small nodules and large tumors. The other is regularizing the surface evolution by applying a Hessian-matrix-based filter that can enhance the vessel structures. Experimental results showed that the proposed method improves sensitivity by over 15% and decreases FP by over 20%, demonstrating that the proposed method is useful for detecting HCC accurately.

With the rapid development and wide application of medical imaging technology, explosive volumes of medical image data are produced every day all over the world. As such, it becomes increasingly challenging to manage and utilize such data effectively and efficiently. In particular, content-based medical image retrieval has been intensively researched in the past decade or so. In this work, we propose a novel approach to content-based medical image retrieval utilizing the co-occurrence of both the texture and the shape features in contrast to most previous algorithms that use purely the texture or the shape feature. Specifically, we propose a novel form of representation for the co-occurrence of the texture and the shape features in an image, i.e., the gray level and edge direction co-occurrence matrix (GLEDCOM). Based on GLEDCOM, we define eleven features forming a feature vector that is used to measure the similarity between images. As a result, it consistently yields outstanding performance on both images rich in texture (e.g., image of brain) and images with dominant smooth regions and sharp edges (e.g., image of bladder). As demonstrated by experiments, the mean precision of retrieval with GLEDCOM algorithm outperforms a set of representative algorithms including the gray level co-occurrence matrix (GLCM) based, the Hu's seven moment invariants (HSMI) based, the uniformity estimation method (UEM) based and the the modified Zernike moments (MZM) based algorithms by 10%-20%.

The challenge in prostate cancer radiotherapy is to deliver the planned dose to the prostate, sparing as much as possible the neighboring organs, namely bladder and rectum. If a lower amount of dose, compared to the prescription, is delivered to the prostate, the risk of failure may increase. Likewise, if higher doses are delivered to the neighboring organs, undesirable side effects may occur. Accurate localization of prostate and organs at risk is therefore a bottleneck in radiotherapy. In recent Image Guided Radiotherapy (IGRT) procedures, an intra-operative Cone Beam CT (CBCT) is used at each session to align the prostate to the planned CT and to maximize the correct dose delivery. Tracking the prostate in these images may allow not only to achieve this goal but also to accurately measure the cumulated dose as the session goes. This work introduces a new method that automatically locates the prostate in CBCT images. The whole method lies in a Bayesian formulation where a multiscale image representation, the local jets, is used as a likelihood function, an the prior knowledge is learned from multiple examples by expert manual delineations. Compared with manual ground truth segmentations, the results showed a Jaccard similarity index of 0.84, and an accuracy of 98%in a set of 4 studies of four patients.

In this paper, we present evaluation results for a novel colonic polyp classification method for use as part of a computed tomographic colonography (CTC) computer-aided detection (CAD) algorithm. Inspired by the interpretative methodology of radiologists using 3D fly-through mode in CTC reading, we have developed an algorithm which utilizes sequences of images (referred to here as videos) for classification of CAD marks. First, we generated an initial list of polyp candidates using an existing CAD system. For each of these candidates, we created a video composed of a series of intraluminal, volume-rendered images focusing on the candidate from multiple viewpoints. These videos illustrated the shape of the polyp candidate and gathered contextual information of diagnostic importance. We calculated the histogram of oriented gradients (HOG) feature on each frame of the video and utilized a support vector machine for classification. We tested our method by analyzing a CTC data set of 50 patients from three medical centers. Our proposed video analysis method for polyp classification showed significantly better performance than an approach using only the 2D CT slice data. The areas under the ROC curve for these methods were 0.88 (95% CI: [0.84, 0.91]) and 0.80 (95% CI: [0.75, 0.84]) respectively (p=0.0005).

Segmentation of the mesenteric vasculature has important applications for evaluation of the small bowel. In particular, it may be useful for small bowel path reconstruction and precise localization of small bowel tumors such as carcinoid. Segmentation of the mesenteric vasculature is very challenging, even for manual labeling, because of the low contrast and tortuosity of the small blood vessels. Many vessel segmentation methods have been proposed. However, most of them are designed for segmenting large vessels. We propose a semi-automated method to extract the mesenteric vasculature on contrast-enhanced abdominal CT scans. First, the internal abdominal region of the body is automatically identified. Second, the major vascular branches are segmented using a multi-linear vessel tracing method. Third, small mesenteric vessels are segmented using multi-view multi-scale vesselness enhancement filters. The method is insensitive to image contrast, variations of vessel shape and small occlusions due to overlapping. The method could automatically detect mesenteric vessels with diameters as small as 1 mm. Compared with the standard-of-reference manually labeled by an expert radiologist, the segmentation accuracy (recall rate) for the whole mesenteric vasculature was 82.3% with a 3.6% false positive rate.

In this paper, we propose a segmentation algorithm by which anatomical landmarks on the pelvis are extracted from computed tomography (CT) images. The landmarks are used to automatically define the anterior (AASA) and posterior acetabular sector angles (PASA) describing the degree of hip misalignment. The center of each femoral head is obtained by searching for the point at which most intensity gradient vectors defined at edge points intersect. The radius of each femoral head is computed by finding the sphere, positioned at the center of the femoral head, for which the normalized sum of gradient vector magnitudes on the sphere surface is maximal. The anterior and posterior corners of each acetabulum are searched for on a curve representing the acetabulum and defined by dynamic programming. The femoral head centers and anterior and posterior corners are used to calculate the AASA and PASA. The algorithm was applied to CT images of 120 normal subjects and the results were compared to ground truth values obtained by manual segmentation. The mean absolute difference (± standard deviation) between the obtained and ground truth values was 1.3 ± 0.3 mm for the femoral head centers and 2.1 ± 1.3 degrees for the acetabular angles.

Osteoarthritis is one of the leading causes of pain and disability worldwide and a major health problem in developed countries due to the gradually aging population. Though the symptoms are easily recognized and described by a patient, it is difficult to assess the level of damage or loss of articular cartilage quantitatively. We present a novel method for fully automated knee cartilage thickness measurement and subsequent assessment of the knee joint. First, the point correspondence between a pre-segmented training bone model is obtained with use of Shape Context based non-rigid surface registration. Then, a single Active Shape Model (ASM) is used to segment both Femur and Tibia bone. The surfaces obtained are processed to extract the Bone-Cartilage Interface (BCI) points, where the proper segmentation of cartilage begins. For this purpose, the cartilage ASM is trained with cartilage edge positions expressed in 1D coordinates at the normals in the BCI points. The whole cartilage model is then constructed from the segmentations obtained in the previous step. An absolute thickness of the segmented cartilage is measured and compared to the mean of all training datasets, giving as a result the relative thickness value. The resulting cartilage structure is visualized and related to the segmented bone. In this way the condition of the cartilage is assessed over the surface. The quality of bone and cartilage segmentation is validated and the Dice's coefficients 0.92 and 0.86 for Femur and Tibia bones and 0.45 and 0.34 for respective cartilages are obtained. The clinical diagnostic relevance of the obtained thickness mapping is being evaluated retrospectively. We hope to validate it prospectively for prediction of clinical outcome the methods require improvements in accuracy and robustness.

To improve the clinical assessment of osteoporotic hip fracture risk, recent computer-aided diagnosis systems explore new approaches to estimate the local trabecular bone quality beyond bone density alone to predict femoral bone strength. In this context, statistical bone mineral density (BMD) features extracted from multi-detector computed tomography (MDCT) images of proximal femur specimens and different function approximations methods were compared in their ability to predict the biomechanical strength. MDCT scans were acquired in 146 proximal femur specimens harvested from human cadavers. The femurs' failure load (FL) was determined through biomechanical testing. An automated volume of interest (VOI)-fitting algorithm was used to define a consistent volume in the femoral head of each specimen. In these VOIs, the trabecular bone was represented by statistical moments of the BMD distribution and by pairwise spatial occurrence of BMD values using the gray-level co-occurrence (GLCM) approach. A linear multi-regression analysis (MultiReg) and a support vector regression algorithm with a linear kernel (SVRlin) were used to predict the FL from the image feature sets. The prediction performance was measured by the root mean square error (RMSE) for each image feature on independent test sets; in addition the coefficient of determination R² was calculated. The best prediction result was obtained with a GLCM feature set using SVRlin, which had the lowest prediction error (RSME = 1.040±0.143, R² = 0.544) and which was significantly lower that the standard approach of using BMD.mean and MultiReg (RSME = 1.093±0.133, R² = 0.490, p<0.0001). The combined sets including BMD.mean and GLCM features had a similar or slightly lower performance than using only GLCM features. The results indicate that the performance of high-dimensional BMD features extracted from MDCT images in predicting the biomechanical strength of proximal femur specimens can be significantly improved by using support vector regression.

Vertebral compression fractures can be caused by even minor trauma in patients with pathological conditions such as osteoporosis, varying greatly in vertebral body location and compression geometry. The location and morphology of the compression injury can guide decision making for treatment modality (vertebroplasty versus surgical fixation), and can be important for pre-surgical planning. We propose a height compass to evaluate the axial plane spatial distribution of compression injury (anterior, posterior, lateral, and central), and distinguish it from physiologic height variations of normal vertebrae. The method includes four steps: spine segmentation and partition, endplate detection, height compass computation and compression fracture evaluation. A height compass is computed for each vertebra, where the vertebral body is partitioned in the axial plane into 17 cells oriented about concentric rings. In the compass structure, a crown-like geometry is produced by three concentric rings which are divided into 8 equal length arcs by rays which are subtended by 8 common central angles. The radius of each ring increases multiplicatively, with resultant structure of a central node and two concentric surrounding bands of cells, each divided into octants. The height value for each octant is calculated and plotted against octants in neighboring vertebrae. The height compass shows intuitive display of the height distribution and can be used to easily identify the fracture regions. Our technique was evaluated on 8 thoraco-abdominal CT scans of patients with reported compression fractures and showed statistically significant differences in height value at the sites of the fractures.

In this paper, we introduce a new local learning based approach and apply it for the well-studied problem of breast cancer diagnosis using BIRADS-based mammographic features. To learn from our clinical dataset the latent relationship between these features and the breast biopsy result, our method first dynamically partitions the whole sample population into multiple sub-population groups through stochastically searching the sample population clustering space. Each encountered clustering scheme in our online searching process is then used to create a certain sample population partition plan. For every resultant sub-population group identified according to a partition plan, our method then trains a dedicated local learner to capture the underlying data relationship. In our study, we adopt the linear logistic regression model as our local learning method's base learner. Such a choice is made both due to the well-understood linear nature of the problem, which is compellingly revealed by a rich body of prior studies, and the computational efficiency of linear logistic regression--the latter feature allows our local learning method to more effectively perform its search in the sample population clustering space. Using a database of 850 biopsy-proven cases, we compared the performance of our method with a large collection of publicly available state-of-the-art machine learning methods and successfully demonstrated its performance advantage with statistical significance.

In this study, a novel mammogram enhancement solution is proposed, aiming to improve the quality of subsequent mass segmentation in mammograms. It has been widely accepted that characteristics of masses are usually hyper-dense or uniform density with respect to its background. Also, their core parts are likely to have high-intensity values while the values of intensity tend to be decreased as the distance to core parts increases. Based on the aforementioned observations, we develop a new and effective mammogram enhancement method by combining local statistical measurements and Sliding Band Filtering (SBF). By effectively combining local statistical measurements and SBF, we are able to improve the contrast of the bright and smooth regions (which represent potential mass regions), as well as, at the same time, the regions where their surrounding gradients are converging to the centers of regions of interest. In this study, 89 mammograms were collected from the public MAIS database (DB) to demonstrate the effectiveness of the proposed enhancement solution in terms of improving mass segmentation. As for a segmentation method, widely used contour-based segmentation approach was employed. The contour-based method in conjunction with the proposed enhancement solution achieved overall detection accuracy of 92.4% with a total of 85 correct cases. On the other hand, without using our enhancement solution, overall detection accuracy of the contour-based method was only 78.3%. In addition, experimental results demonstrated the feasibility of our enhancement solution for the purpose of improving detection accuracy on mammograms containing dense parenchymal patterns.

In the paper, we propose a novel scheme for breast mass segmentation in mammography, which is based on visual perception and consists of two steps. Firstly, radiologists' eye-gazing data is recorded by the eye-tracker during reading and then clustered with a density-based spatial clustering of applications with noise (DBSCAN) algorithm to achieve seeds locating radiologists' regions of interest (ROIs). The seeds-based region growing (SBRG) algorithm is applied to buckle ROIs containing suspicious lesions. Secondly, in order to achieve fine lesion contour as final result, the ROIs are segmented with a multi-scale mass segmentation approach using active contour models. The result of applying the proposed method to the mammograms from both DDSM and Zhejiang Cancer Hospital shows that the achieved average of overlap rate is 0.5915 and the achieved average of misclassification rate is 0.6342. The innovative point of the proposed approach is to introduce visual perception into breast mass segmentation, which makes the result of mass segmentation meet radiologists' subjective demand.

We present a method using statistical measures of the orientation of texture to characterize and detect architectural distortion in prior mammograms of interval-cancer cases. Based on the orientation field, obtained by the application of a bank of Gabor filters to mammographic images, two types of co-occurrence matrices were derived to estimate the joint occurrence of the angles of oriented structures. For each of the matrices, Haralick's 14 texture features were computed. From a total of 106 prior mammograms of 56 interval-cancer cases and 52 mammograms of 13 normal cases, 4,224 regions of interest (ROIs) were automatically obtained by applying Gabor filters and phase portrait analysis. For each ROI, statistical features were computed using the angle co-occurrence matrices. The performance of the features in the detection of architectural distortion was analyzed and compared with that of Haralick's features computed using the gray-level co-occurrence matrices of the ROIs. Using logistic regression for feature selection, an artificial neural network for classification, and the leave-one-image-out approach for cross-validation, the best result achieved was 0.77 in terms of the area under the receiver operating characteristic (ROC) curve. Analysis of the free-response ROC curve yielded a sensitivity of 80% at 5.4 false positives per image.

This paper presents a Computer-Aided Diagnosis (CAD) system for mammograms, which is based on complex networks to shape boundary characterization of mass in mammograms, suggesting a "second opinion" to the health specialist. A region of interest (the mass) is automatically segmented using an improved algorithm based on EM/MPM and the shape is modeled into a scale-free complex network. Topological measurements of the resulting network are used to compose the shape descriptors. The experiments comparing the complex network approach with other traditional descriptors, in detecting breast cancer in mammograms, show that the proposed approach accomplish the best values of accuracy. Hence, the results indicate that complex networks are wellsuited to characterize mammograms.

Breast cancer is one of the most common malignant tumors in women. In mammogram retrieval system, the query mass is ambiguity and difficult to be described because in which the lesion and the normal tissue are physically adjacent. If the query mass can be processed as an image bag, then the ambiguity can be tackled by multi-instance learning (MIL) techniques. In this paper, we presented a preliminary study of MIL for mass retrieval in digitized mammograms, and proposed three image bag generators named J-Bag, A-Bag and K-Bag, respectively. Diverse Density (DD), EM-DD and BP-MIP were applied as MIL algorithms for mass retrieval. Experimental study was carried out on DDSM database and another database in which images were collected from the Zhejiang Cancer Hospital in China. Preliminary experiments showed that the MIL techniques can be applied to the problem of mass retrieval in digitized mammograms and the proposed bag generators A-Bag and K-Bag can achieve more efficient results than the existing bag generator SBN.

Nuclei counting in epithelial cells is an indication for tumor proliferation rate which is useful to rank tumors and select an appropriate treatment schedule for the patient. However, due to the high interand intra- observer variability in nuclei counting, pathologists seek a deterministic proliferation rate estimate. Histology tissue contains epithelial and stromal cells. However, nuclei counting is clinically restricted to epithelial cells because stromal cells do not become cancerous themselves since they remain genetically normal. Counting nuclei existing within the stromal tissue is one of the major causes of the proliferation rate non-deterministic estimation. Digitally removing stromal tissue will eliminate a major cause in pathologist counting variability and bring the clinical pathologist a major step closer toward a deterministic proliferation rate estimation. To that end, we propose a computer aided diagnosis (CAD) system for eliminating stromal cells from digital histology images based on the local binary patterns, entropy measurement, and statistical analysis. We validate our CAD system on a set of fifty Ki-67-stained histology images. Ki-67-stained histology images are among the clinically approved methods for proliferation rate estimation. To test our CAD system, we prove that the manual proliferation rate estimation performed by the expert pathologist does not change before and after stromal removal. Thus, stromal removal does not affect the expert pathologist estimation clinical decision. Hence, the successful elimination of the stromal area highly reduces the false positive nuclei which are the major confusing cause for the less experienced pathologists and thus accounts for the non-determinism in the proliferation rate estimation. Our experimental setting shows statistical insignificance (paired student t-test shows ρ = 0.74) in the manual nuclei counting before and after our automated stromal removal. This means that the clinical decision of the expert pathologist is not affected by our CAD system which is what we want to prove. However, the usage of our CAD system substantially account for the reduced inter- and intra- proliferation rate estimation variability and especially for less-experienced pathologists.

The presence of axillary lymph node metastases is the most important prognostic factor in breast cancer and can influence the selection of adjuvant therapy, both chemotherapy and radiotherapy. In this work we present a set of kinetic statistics derived from DCE-MRI for predicting axillary node status. Breast DCE-MRI images from 69 women with known nodal status were analyzed retrospectively under HIPAA and IRB approval. Axillary lymph nodes were positive in 12 patients while 57 patients had no axillary lymph node involvement. Kinetic curves for each pixel were computed and a pixel-wise map of time-to-peak (TTP) was obtained. Pixels were first partitioned according to the similarity of their kinetic behavior, based on TTP values. For every kinetic curve, the following pixel-wise features were computed: peak enhancement (PE), wash-in-slope (WIS), wash-out-slope (WOS). Partition-wise statistics for every feature map were calculated, resulting in a total of 21 kinetic statistic features. ANOVA analysis was done to select features that differ significantly between node positive and node negative women. Using the computed kinetic statistic features a leave-one-out SVM classifier was learned that performs with AUC=0.77 under the ROC curve, outperforming the conventional kinetic measures, including maximum peak enhancement (MPE) and signal enhancement ratio (SER), (AUCs of 0.61 and 0.57 respectively). These findings suggest that our DCE-MRI kinetic statistic features can be used to improve the prediction of axillary node status in breast cancer patients. Such features could ultimately be used as imaging biomarkers to guide personalized treatment choices for women diagnosed with breast cancer.

This paper presents a novel scheme for mass segmentation in digitized mammograms, which is based on Graph Cuts algorithm and multi-scale analysis. The multi-scale method can segment mammographic images with a stepwise process from global to local segmentation by iterating Graph Cuts. To improve the segmentation efficiency and robustness, the watershed transform is used for pre-segmentation of the image to produce a region adjacency graph for the following optimization steps. Besides, this paper proposes a strategy of increasing smoothness energy term step by step in the Markov Random Field (MRF) image segmentation module, so as to improve the efficiency effectively in the mass segmentation. The new segmentation strategy effectively improves segmentation performance with less influences of image noise level. The experimental results demonstrate that the proposed method achieves a better performance in accuracy and robustness than conventional ones.

Breast cancer is one of the main causes of death among women in occidental countries. In the last years, screening mammography has been established worldwide for early detection of breast cancer, and computer-aided diagnostics (CAD) is being developed to assist physicians reading mammograms. A promising method for CAD is content-based image retrieval (CBIR). Recently, we have developed a classification scheme of suspicious tissue pattern based on the support vector machine (SVM). In this paper, we continue moving towards automatic CAD of screening mammography. The experiments are based on in total 10,509 radiographs that have been collected from different sources. From this, 3,375 images are provided with one and 430 radiographs with more than one chain code annotation of cancerous regions. In different experiments, this data is divided into 12 and 20 classes, distinguishing between four categories of tissue density, three categories of pathology and in the 20 class problem two categories of different types of lesions. Balancing the number of images in each class yields 233 and 45 images remaining in each of the 12 and 20 classes, respectively. Using a two-dimensional principal component analysis, features are extracted from small patches of 128 x 128 pixels and classified by means of a SVM. Overall, the accuracy of the raw classification was 61.6 % and 52.1 % for the 12 and the 20 class problem, respectively. The confusion matrices are assessed for detailed analysis. Furthermore, an implementation of a SVM-based CBIR system for CADx in screening mammography is presented. In conclusion, with a smarter patch extraction, the CBIR approach might reach precision rates that are helpful for the physicians. This, however, needs more comprehensive evaluation on clinical data.

We are developing a Decision Tree Content-Based Image Retrieval (DTCBIR) CADx scheme to assist radiologists in characterization of breast masses on ultrasound (US) images. Three DTCBIR configurations, including decision tree with boosting (DTb), decision tree with full leaf features (DTL), and decision tree with selected leaf features (DTLs) were compared. For DTb, the features of a query mass were combined first into a merged feature score and then masses with similar scores were retrieved. For DTL and DTLs, similar masses were retrieved based on the Euclidean distance between the feature vector of the query and those of the selected references. For each DTCBIR configuration, we investigated the use of the full feature set and the subset of features selected by the stepwise linear discriminant analysis (LDA) and simplex optimization method, resulting in six retrieval methods. Among the six methods, we selected five, DTb-lda, DTL-lda, DTb-full, DTL-full and DTLs-full, for the observer study. For a query mass, three most similar masses were retrieved with each method and were presented to the radiologists in random order. Three MQSA radiologists rated the similarity between the query mass and the computer-retrieved masses using a ninepoint similarity scale (1=very dissimilar, 9=very similar). For DTb-lda, DTL-lda, DTb-full, DTL-full and DTLs-full, the average A_z values were 0.90±0.03, 0.85±0.04, 0.87±0.04, 0.79±0.05 and 0.71±0.06, respectively, and the average similarity ratings were 5.00, 5.41, 4.96, 5.33 and 5.13, respectively. Although the DTb measures had the best classification performance among the DTCBIRs studied, and DTLs had the worst performance, DTLs-full obtained higher similarity ratings than the DTb measures.

In this paper we propose a new method to segment a breast image into several regions. Tumor detection region is constrained to the region only in glandular tissue because the tumors usually occur at glandular tissue in the breast anatomy. We extract texture feature for each point and classify them as several layers using a random forest classifier. Classified points are merged into a large region and small regions are removed by postprocessing. The accuracy of glandular tissue detection rate was about 90%. We applied the conventional tumor detection method in this segmented glandular tissue. After several tests we obtained that tumor detection accuracy improved for 14% and detection time was also reduced. With this method, we can achieve the improvement both on tumor detection accuracy and on the processing time.

Breast cancer is the second cause of women death and the most diagnosed female cancer in the US. Proliferation rate estimation (PRE) is one of the prognostic indicators that guide the treatment protocols and it is clinically performed from Ki-67 histopathology images. Automating PRE substantially increases the efficiency of the pathologists. Moreover, presenting a deterministic and reproducible proliferation rate value is crucial to reduce inter-observer variability. To that end, we propose a fully automated CAD system for PRE from the Ki-67 histopathology images. This CAD system is based on a model of three steps: image pre-processing, image clustering, and nuclei segmentation and counting that are finally followed by PRE. The first step is based on customized color modification and color-space transformation. Then, image pixels are clustered by K-Means depending on the features extracted from the images derived from the first step. Finally, nuclei are segmented and counted using global thresholding, mathematical morphology and connected component analysis. Our experimental results on fifty Ki-67-stained histopathology images show a significant agreement between our CAD's automated PRE and the gold standard's one, where the latter is an average between two observers' estimates. The Paired T-Test, for the automated and manual estimates, shows ρ = 0.86, 0.45, 0.8 for the brown nuclei count, blue nuclei count, and proliferation rate, respectively. Thus, our proposed CAD system is as reliable as the pathologist estimating the proliferation rate. Yet, its estimate is reproducible.

We investigated the feasibility of using multiresolution Local Binary Pattern (LBP) texture analysis to reduce falsepositive (FP) detection in a computerized mass detection framework. A new and novel approach for extracting LBP features is devised to differentiate masses and normal breast tissue on mammograms. In particular, to characterize the LBP texture patterns of the boundaries of masses, as well as to preserve the spatial structure pattern of the masses, two individual LBP texture patterns are then extracted from the core region and the ribbon region of pixels of the respective ROI regions, respectively. These two texture patterns are combined to produce the so-called multiresolution LBP feature of a given ROI. The proposed LBP texture analysis of the information in mass core region and its margin has clearly proven to be significant and is not sensitive to the precise location of the boundaries of masses. In this study, 89 mammograms were collected from the public MAIS database (DB). To perform a more realistic assessment of FP reduction process, the LBP texture analysis was applied directly to a total of 1,693 regions of interest (ROIs) automatically segmented by computer algorithm. Support Vector Machine (SVM) was applied for the classification of mass ROIs from ROIs containing normal tissue. Receiver Operating Characteristic (ROC) analysis was conducted to evaluate the classification accuracy and its improvement using multiresolution LBP features. With multiresolution LBP features, the classifier achieved an average area under the ROC curve, , z A of 0.956 during testing. In addition, the proposed LBP features outperform other state-of-the-arts features designed for false positive reduction.

Dedicated breast CT (bCT) is an emerging technology that produces 3D images of the breast, thus allowing radiologists to detect and evaluate breast lesions in 3D. However, assessing potential cancers in the bCT volume can prove time consuming and difficult. Thus, we are developing automated 3D lesion segmentation methods to aid in the interpretation of bCT images. Based on previous studies using a 3D radial-gradient index (RGI) method [1], we are investigating whether 3D active contour segmentation can be applied in 3D to capture additional details of the lesion margin. Our data set includes 40 contract-enhanced bCT scans. Based on a radiologist-marked lesion center of each mass, an initial RGI contour is obtained that serves as the input to an active contour segmentation method. In this study, active contour level set segmentation, an iterative segmentation technique, is extended to 3D. Three stopping criteria are compared, based on 1) the change of volume (ΔV/V), 2) the mean value of the increased volume at each iteratin (dμ/dt), and 3) the changing rate of intensity inside and outside the lesion (Δv_w). Lesion segmentation was evaluated by determining the overlap ratio between computer-determined segmentations and manually-drawn lesion outlines. For a given lesion, the overlap ratio was averaged across coronal, sagittal, and axial planes. The average overlap ratios for the three stopping criteria were found to be 0.66 (ΔV/V), 0.68 (dμ/dt), 0.69 (Δv_w).

DBT is one of the promising imaging modalities that may improve the sensitivity and specificity for breast cancer detection. We are developing a computer-aided detection (CADe) system for clustered microcalcifications (MC) in DBT. A data set of two-view DBTs from 42 breasts was collected with a GE prototype system. We investigated a 2D approach to MC detection using projection view (PV) images rather than reconstructed 3D DBT volume. Our 2D approach consisted of two major stages: 1) detecting individual MC candidates on each PV, and 2) correlating the MC candidates from the different PVs and detecting clusters in the breast volume. With the MC candidates detected by prescreening on PVs, a trained multi-channel (MCH) filter bank was used to extract signal response from each MC candidate. A ray-tracing process was performed to fuse the MCH responses and localize the MC candidates in 3D using the geometrical information of the DBT system. Potential MC clusters were then identified by dynamic clustering of the MCs in 3D. A two-fold cross-validation method was used to train and test the CADe system. The detection performance of clustered MCs was assessed by free receiver operating characteristic (FROC) analysis. It was found that the CADe system achieved a case-based sensitivity of 90% at an average false positive rate of 2.1 clusters per DBT volume. Our study demonstrated that the CADe system using 2D MCH filter bank is promising for detection of clustered MCs in DBT.

Dedicated cone-beam breast CT (bCT) scanners have been developed as a potential alternative imaging modality to conventional X-ray mammography in breast cancer diagnosis. As with other modalities, quantitative imaging (QI) analysis can potentially be utilized as a tool to extract useful numeric information concerning diagnosed lesions from high quality 3D tomographic data sets. In this work, preliminary QI analysis was done by designing and implementing a computer-aided diagnosis (CADx) system consisting of image preprocessing, object(s) of interest (i.e. masses, microcalcifications) segmentation, structural analysis of the segmented object(s), and finally classification into benign or malignant disease. Image sets were acquired from bCT patient scans with diagnosed lesions. Iterative watershed segmentation (IWS), a hybridization of the watershed method using observer-set markers and a gradient vector flow (GVF) approach, was used as the lesion segmentation method in 3D. Eight morphologic parameters and six texture features based on gray level co-occurrence matrix (GLCM) calculations were obtained per segmented lesion and combined into multi-dimensional feature input data vectors. Artificial neural network (ANN) classifiers were used by performing cross validation and network parameter optimization to maximize area under the curve (AUC) values of the resulting receiver-operating characteristic (ROC) curves. Within these ANNs, biopsy-proven diagnoses of malignant and benign lesions were recorded as target data while the feature vectors were saved as raw input data. With the image data separated into post-contrast (n = 55) and pre-contrast sets (n = 39), a maximum AUC of 0.70 ± 0.02 and 0.80 ± 0.02 were achieved, respectively, for each data set after ANN application.

In this paper we propose a novel approach based on multi-stage random forests to address problems faced by traditional vessel segmentation algorithms on account of image artifacts such as stitches organ shadows etc.. Our approach consists of collecting a very large number of training data consisting of positive and negative examples of valid seed points. The method makes use of a 14x14 window around a putative seed point. For this window three types of feature vectors are computed viz. vesselness, eigenvalue and a novel effective margin feature. A random forest RF is trained for each of the feature vectors. At run time the three RFs are applied in succession to a putative seed point generated by a naiive vessel detection algorithm based on vesselness. Our approach will prune this set of putative seed points to correctly identify true seed points thereby avoiding false positives. We demonstrate the effectiveness of our algorithm on a large dataset of angio images.

Changes in myocardial function signatures such as wall motion and thickening are typically computed separately from myocardial perfusion SPECT (MPS) stress and rest studies to assess for stress-induced function abnormalities. The standard approach may suffer from the variability in contour placements and image orientation when subtle changes between stress and rest scans in motion and thickening are being evaluated. We have developed a new measure of regional change of function signature (motion and thickening) computed directly from registered stress and rest gated MPS data. In our novel approach, endocardial surfaces at the end-diastolic and end-systolic frames for stress and rest studies were registered by matching ventricular surfaces. Furthermore, we propose a new global registration method based on finding the optimal rotation for myocardial best ellipsoid fit to minimize the indexing disparities between two surfaces between stress and rest studies. Myocardial stress-rest function changes were computed and normal limits of change were determined as the mean and standard deviation of the training set for each polar sample. Normal limits were utilized to quantify the stress-rest function change for each polar map sample and the accumulated quantified function signature values were used for abnormality assessments in territorial regions. To evaluate the effectiveness of our novel method, we examined the agreements of our results against visual scores for motion change on vessel territorial regions obtained by human experts on a test group with 623 cases and were able to show that our detection method has a improved sensitivity on per vessel territory basis, compared to those obtained by human experts utilizing gated MPS data.

Carotid atherosclerosis is a major cause of stroke, a leading cause of death and disability. In this paper, we develop and evaluate a new segmentation method for outlining both lumen and adventitia (inner and outer walls) of common carotid artery (CCA) from three-dimensional ultrasound (3D US) images for carotid atherosclerosis diagnosis and evaluation. The data set consists of sixty-eight, 17× 2× 2, 3D US volume data acquired from the left and right carotid arteries of seventeen patients (eight treated with 80mg atorvastain and nine with placebo), who had carotid stenosis of 60% or more, at baseline and after three months of treatment. We investigate the use of Active Shape Models (ASMs) to segment CCA inner and outer walls after statin therapy. The proposed method was evaluated with respect to expert manually outlined boundaries as a surrogate for ground truth. For the lumen and adventitia segmentations, respectively, the algorithm yielded Dice Similarity Coefficient (DSC) of 93.6%± 2.6%, 91.8%± 3.5%, mean absolute distances (MAD) of 0.28± 0.17mm and 0.34 ± 0.19mm, maximum absolute distances (MAXD) of 0.87 ± 0.37mm and 0.74 ± 0.49mm. The proposed algorithm took 4.4 ± 0.6min to segment a single 3D US images, compared to 11.7±1.2min for manual segmentation. Therefore, the method would promote the translation of carotid 3D US to clinical care for the fast, safety and economical monitoring of the atherosclerotic disease progression and regression during therapy.

Inter- and intra- observer variability is a problem often faced when an expert or observer is tasked with assessing the severity of a disease. This issue is keenly felt in coronary calcium scoring of patients suffering from atherosclerosis where in clinical practice, the observer must identify firstly the presence, followed by the location of candidate calcified plaques found within the coronary arteries that may prevent oxygenated blood flow to the heart muscle. However, it can be difficult for a human observer to differentiate calcified plaques that are located in the coronary arteries from those found in surrounding anatomy such as the mitral valve or pericardium. In addition to the benefits to scoring accuracy, the use of fast, low dose multi-slice CT imaging to perform the cardiac scan is capable of acquiring the entire heart within a single breath hold. Thus exposing the patient to lower radiation dose, which for a progressive disease such as atherosclerosis where multiple scans may be required, is beneficial to their health. Presented here is a fully automated method for calcium scoring using both the traditional Agatston method, as well as the volume scoring method. Elimination of the unwanted regions of the cardiac image slices such as lungs, ribs, and vertebrae is carried out using adaptive heart isolation. Such regions cannot contain calcified plaques but can be of a similar intensity and their removal will aid detection. Removal of both the ascending and descending aortas, as they contain clinical insignificant plaques, is necessary before the final calcium scores are calculated and examined against ground truth scores of three averaged expert observer results. The results presented here are intended to show the feasibility and requirement for an automated scoring method to reduce the subjectivity and reproducibility error inherent with manual clinical calcium scoring.

Minimally invasive catheter ablation of electric foci, performed in electrophysiology labs, is an attractive treatment option for atrial fibrillation (AF) - in particular if drug therapy is no longer effective or tolerated. There are different strategies to eliminate the electric foci inducing the arrhythmia. Independent of the particular strategy, it is essential to place transmural lesions. The impact of catheter contact force on the generated lesion quality has been investigated recently, and first results are promising. There are different approaches to measure catheter-tissue contact. Besides traditional haptic feedback, there are new technologies either relying on catheter tip-to-tissue contact force or on local impedance measurements at the tip of the catheter. In this paper, we present a novel tool for post-procedural ablation point evaluation and visualization of contact force characteristics. Our method is based on localizing ablation points set during AF ablation procedures. The 3-D point positions are stored together with lesion specific catheter contact force (CF) values recorded during the ablation. The force records are mapped to the spatial 3-D positions, where the energy has been applied. The tracked positions of the ablation points can be further used to generate a 3-D mesh model of the left atrium (LA). Since our approach facilitates visualization of different force characteristics for post-procedural evaluation and verification, it has the potential to improve outcome by highlighting areas where lesion quality may be less than desired.

To identify asymptomatic patients is the challenging task and the essential first step in diagnosis. Findings of dental panoramic radiographs include not only dental conditions but also radiographic signs that are suggestive of possible systemic diseases such as osteoporosis, arteriosclerosis, and maxillary sinusitis. Detection of such signs on panoramic radiographs has a potential to provide supplemental benefits for patients. However, it is not easy for general dental practitioners to pay careful attention to such signs. We addressed the development of a computer-aided detection (CAD) system that detects radiographic signs of pathology on panoramic images, and the design of the framework of new screening pathway by cooperation of dentists and our CAD system. The performance evaluation of our CAD system showed the sensitivity and specificity in the identification of osteoporotic patients were 92.6 % and 100 %, respectively, and those of the maxillary sinus abnormality were 89.6 % and 73.6 %, respectively. The detection rate of carotid artery calcifications that suggests the need for further medical evaluation was approximately 93.6 % with 4.4 false-positives per image. To validate the utility of the new screening pathway, preliminary clinical trials by using our CAD system were conducted. To date, 223 panoramic images were processed and 4 asymptomatic patients with suspected osteoporosis, 7 asymptomatic patients with suspected calcifications, and 40 asymptomatic patients with suspected maxillary sinusitis were detected in our initial trial. It was suggested that our new screening pathway could be useful to identify asymptomatic patients with systemic diseases.

Findings of dental panoramic radiographs (DPRs) have shown that the mandibular cortical thickness (MCT) was significantly correlated with osteoporosis. Identifying asymptomatic patients with osteoporosis through dental examinations may bring a supplemental benefit for the patients. However, most of the DPRs are used for only diagnosing dental conditions by dentists in their routine clinical work. The aim of this study was to develop a computeraided diagnosis scheme that automatically measures MCT to assist dentists in screening osteoporosis. First, the inferior border of mandibular bone was detected by use of an active contour method. Second, the locations of mental foramina were estimated on the basis of the inferior border of mandibular bone. Finally, MCT was measured on the basis of the grayscale profile analysis. One hundred DPRs were used to evaluate our proposed scheme. Experimental results showed that the sensitivity and specificity for identifying osteoporotic patients were 92.6 % and 100 %, respectively. We conducted multiclinic trials, in which 223 cases have been obtained and processed in about a month. Our scheme succeeded in detecting all cases of suspected osteoporosis. Therefore, our scheme may have a potential to identify osteoporotic patients at an early stage.

The apical root regions play an important role in analysis and diagnosis of many oral diseases. Automatic detection of such regions is consequently the first step toward computer-aided diagnosis of these diseases. In this paper we propose an automatic method for periapical root region detection by using the state-of-theart machine learning approaches. Specifically, we have adapted the AdaBoost classifier for apical root detection. One challenge in the task is the lack of training cases especially for diseased ones. To handle this problem, we boost the training set by including more root regions that are close to the annotated ones and decompose the original images to randomly generate negative samples. Based on these training samples, the Adaboost algorithm in combination with Haar wavelets is utilized in this task to train an apical root detector. The learned detector usually generates a large amount of true and false positives. In order to reduce the number of false positives, a confidence score for each candidate detection result is calculated for further purification. We first merge the detected regions by combining tightly overlapped detected candidate regions and then we use the confidence scores from the Adaboost detector to eliminate the false positives. The proposed method is evaluated on a dataset containing 39 annotated digitized oral X-Ray images from 21 patients. The experimental results show that our approach can achieve promising detection accuracy.

Despite major improvements in dental healthcare and technology, dental caries remains one of the most prevalent chronic diseases of modern society. The initial stages of dental caries are characterized by demineralization of enamel crystals, commonly known as white spots, which are difficult to diagnose. Near-infrared (NIR) hyperspectral imaging is a new promising technique for early detection of demineralization which can classify healthy and pathological dental tissues. However, due to non-ideal illumination of the tooth surface the hyperspectral images can exhibit specular reflections, in particular around the edges and the ridges of the teeth. These reflections significantly affect the performance of automated classification and visualization methods. Cross polarized imaging setup can effectively remove the specular reflections, however is due to the complexity and other imaging setup limitations not always possible. In this paper, we propose an alternative approach based on modeling the specular reflections of hard dental tissues, which significantly improves the classification accuracy in the presence of specular reflections. The method was evaluated on five extracted human teeth with corresponding gold standard for 6 different healthy and pathological hard dental tissues including enamel, dentin, calculus, dentin caries, enamel caries and demineralized regions. Principal component analysis (PCA) was used for multivariate local modeling of healthy and pathological dental tissues. The classification was performed by employing multiple discriminant analysis. Based on the obtained results we believe the proposed method can be considered as an effective alternative to the complex cross polarized imaging setups.

Retinal images are long-accepted clinical diagnostic method for ocular diseases. Of late, automated assessment of retinal images has proven to be a useful adjunct in clinical decision support systems. In this paper, we propose a retinal image registration method, which combine retinal image enhancement and non-rigid image registration methods, for longitudinal retinal image alignment. A further illumination correction and gray value matching methods are applied for the longitudinal image comparison and subtraction. The solution can enhance the assessment of longitudinal changes of retinal images and image subtraction in a clinical application system. The performance of the proposed solution has been tested on longitudinal retinal images. Preliminary results have demonstrated the accuracy and robustness of the solutions and their potential application in a clinical environment.

We are developing a computer-aided detection (CAD) system to assist radiologists in pulmonary embolism (PE) detection in computed tomographic pulmonary angiography (CTPA). Automatic segmentation and tracking of pulmonary vessels is a fundamental step to define the search space for PE detection. For automated tracking of pulmonary arteries, it is important to accurately identify the seed points to track the left and right pulmonary vessel trees. In this study, we developed an automatic seed point identification and pulmonary main artery (PMA) segmentation method. The seed point was derived from the bifurcation region where the pulmonary trunk artery splits into the left and right. A 3D recursive optimal path finding method (RPF) was developed to find the paths from the bifurcation point to the end of the left and right PMAs. The PMAs were finally extracted along the PMA paths using morphological operation. Two and 18 CTPA cases was used for training and testing, respectively. A set of points in the central luminal space of the PMA were manually marked as the "reference standard" by two experienced chest radiologists using a computer interface. A total of 3870 were marked in the test set. A voxel located on the computer-identified paths of the PMA was counted as a true PMA voxel when its distance to the closest reference standard point is within a threshold. Our results show that 95.6% (17681/18502) and 88.8% (16439/18502) of computer identified PMA path points were within a distance of 10 mm and 8 mm to the closest reference point, respectively, and 100% (18/18) of the seed points were detected in the bifurcation region. 2.7% (104/3870) of the reference standard points were not contained in the computer segmented vessels and counted as false negative points.

Radiologists are adept at recognizing the appearance of lung parenchymal abnormalities in CT scans. However, the inconsistent differential diagnosis, due to subjective aggregation, mandates supervised classification. Towards optimizing Emphysema classification, we introduce a physician-in-the-loop feedback approach in order to minimize uncertainty in the selected training samples. Using multi-view inductive learning with the training samples, an ensemble of Support Vector Machine (SVM) models, each based on a specific pair-wise dissimilarity metric, was constructed in less than six seconds. In the active relearning phase, the ensemble-expert label conflicts were resolved by an expert. This just-in-time feedback with unoptimized SVMs yielded 15% increase in classification accuracy and 25% reduction in the number of support vectors. The generality of relearning was assessed in the optimized parameter space of six different classifiers across seven dissimilarity metrics. The resultant average accuracy improved to 21%. The co-operative feedback method proposed here could enhance both diagnostic and staging throughput efficiency in chest radiology practice.

Diseased airways have been known for several years as a possible contributing factor to airflow limitation in Chronic Obstructive Pulmonary Diseases (COPD). Quantification of disease severity through the evaluation of airway dimensions - wall thickness and lumen diameter - has gained increased attention, thanks to the availability of multi-slice computed tomography (CT). Novel approaches have focused on automated methods of measurement as a faster and more objective means that the visual assessment routinely employed in the clinic. Since the Full-Width Half-Maximum (FWHM) method of airway measurement was introduced two decades ago [1], several new techniques for quantifying airways have been detailed in the literature, but no approach has truly become a standard for such analysis. Our own research group has presented two alternative approaches for determining airway dimensions, one involving a minimum path and the other active contours [2, 3]. With an increasing number of techniques dedicated to the same goal, we decided to take a step back and analyze the differences of these methods. We consequently put to the test our two methods of analysis and the FWHM approach. We first measured a set of 5 airways from a phantom of known dimensions. Then we compared measurements from the three methods to those of two independent readers, performed on 35 airways in 5 patients. We elaborate on the differences of each approach and suggest conclusions on which could be defined as the best one.

A bone opacity suppressed technique using shape-index processing approach has been developed for frontal chest radiography. The image function preserves original lung image textures but equalizing the image contrast of the lungs as a part of post-processing. To determine the benefit of this computerized processing, particular on the investigation of the effect of the bone opacity removal, we conducted a reader study where radiologists read standard chest radiograph alone (unaided) followed by bone opacity suppressed image (aided). Posterioranterior (PA) standard chest radiographs in 368 subjects (122 had confirmed lung cancer) were used for this study. Fifteen Board Certified radiologists participated in the reader study. Each radiologist interpreted the standard image and then the bone suppressed image. Each reader recorded the location of the most suspicious nodule, if any, their level of suspicion and recommendation for clinical action. Detailed analyses were performed to evaluate the observers' performance by tabulating changes of nodule detection inclusive of false-negative turned to true-positive (FN->TP), true-positive turned to false-negative (TP->FN), false-positive turned to turn-negative (FP->TN), and turn-negative turned to false-positive (TN- >FP). Our results indicated that changing rates of FN->TP was 12.35%, TP->FN was 1.37%, FP->TN was 1.14%, and TN->FP was 4.82%, respectively. We also found that 81.85% of the FN->TP events occurred at nodules significantly covered by the rib (50% or more area overlapped with bone opacity). Two major situations caused TP->FN events: (1) other nodule like areas were also enhanced and (2) non-solid nodules were well preserved but less suspicious with the contract equalization.

We propose an automatic segmentation of Ground Glass Opacity (GGO) nodules on chest CT images by histogram modeling and local contrast. First, optimal volume circumscribing a nodule is calculated by clicking inside of GGO nodule. To remove noises while preserving a nodule boundary, anisotropic diffusion filtering is applied to the optimal volume. Second, for deciding an appropriate threshold value of GGO nodule, histogram modeling is performed by Gaussian Mixture Modeling (GMM) with three components such as lung parenchyma, nodule, and chest wall or vessels. Third, the attached chest wall and vessels are separated from the GGO nodules by maximum curvature points linking and morphological erosion with adaptive circular mask. Fourth, initial boundary of GGO nodule is refined using local contrast information. Experimental results show that attached neighbor structures are well separated from GGO nodules while missed GGO region is refined. The proposed segmentation method can be used for measurement of the growth rate of nodule and the proportion of solid portion inside nodule.