This PDF file contains the front matter associated with SPIE Proceedings Volume 6514, including the Title Page, Copyright information, Table of Contents, Introduction (if any), and the Conference Committee listing.

Breast cancer is the leading cancer among American women. The current lifetime risk of developing breast cancer is 13.4% (one in seven). Mammography is the most effective technology presently available for breast cancer screening. With digital mammograms computer-aided detection (CAD) has proven to be a useful tool for radiologists. In this paper, we focus on mass detection that is a common category of breast cancers relative to calcification and architecture distortion. We propose a new mass detection algorithm utilizing Gabor filters, termed as "Gabor Mass Detection" (GMD). There are three steps in the GMD algorithm, (1) preprocessing, (2) generating alarms and (3) classification (reducing false alarms). Down-sampling, quantization, denoising and enhancement are done in the preprocessing step. Then a total of 30 Gabor filtered images (along 6 bands by 5 orientations) are produced. Alarm segments are generated by thresholding four Gabor images of full orientations (Stage-I classification) with image-dependent thresholds computed via histogram analysis. Next a set of edge histogram descriptors (EHD) are extracted from 24 Gabor images (6 by 4) that will be used for Stage-II classification. After clustering EHD features with fuzzy C-means clustering method, a k-nearest neighbor classifier is used to reduce the number of false alarms. We initially analyzed 431 digitized mammograms (159 normal images vs. 272 cancerous images, from the DDSM project, University of South Florida) with the proposed GMD algorithm. And a ten-fold cross validation was used for testing the GMD algorithm upon the available data. The GMD performance is as follows: sensitivity (true positive rate) = 0.88 at false positives per image (FPI) = 1.25, and the area under the ROC curve = 0.83. The overall performance of the GMD algorithm is satisfactory and the accuracy of locating masses (highlighting the boundaries of suspicious areas) is relatively high. Furthermore, the GMD algorithm can successfully detect early-stage (with small values of Assessment & low Subtlety) malignant masses. In addition, Gabor filtered images are used in both stages of classifications, which greatly simplifies the GMD algorithm.

Mutual information is a popular intensity-based image similarity measure mainly used in image registration. This measure has been also very successful as the similarity metric in our knowledge-based computer-assisted detection (CADe) system for the detection of masses in screening mammograms. Our CADe system is designed to assess a new, query case based on its similarity with known cases stored in the knowledge database. However, intensity-based mutual information captures only relationships between the gray level values of corresponding pixels. This study presents a novel advancement of our CADe system by incorporating neighborhood textural information when estimating the mutual information of two images. Specifically, an entropy filter is applied to the images, effectively replacing each image pixel value with its neighborhood entropy. This pixel-based entropy is a localized measure of image texture. Then, the information-theoretic CAD system is asked to make a decision regarding the query case using the texture-based mutual information similarity metric. The entropy-based image enhancement and MI-based decision making processes are repeated at different neighborhood scales. Finally, an artificial network merges intensity-based and texture-based decisions to investigate possible improvements in mass detection performance. Given a database of 1,820 regions of interest (ROIs) extracted from screening mammograms (901 depicting a biopsy-proven mass and 919 depicting normal parenchyma) and a leave-one out sampling scheme, the study showed that our CADe system achieves an ROC area of 0.87±0.01 using the intensity-based ROC. The ROC performance for the texture-based CADe system ranges from 0.69±0.01 to 0.83±0.01 depending on the scale of analysis. The synergistic approach of the ANN using both intensity-based and texture-based information resulted in statistically significantly better performance with an ROC area index of 0.93±0.01.

The study investigates the significance of wavelet-based and MPEG-7 homogeneous textural features in an attempt to improve the specificity of an in-house CAD system for the detection of masses in screening mammograms. The detection scheme has been presented before and it relies on the concept of morphologic concentric layer (MCL) analysis to identify suspicious locations in a mammogram. The locations were deemed suspicious due to their morphology; especially an increased activity of iso-intensity layers around these locations. On a set of 270 mammographic images, the MCL detection scheme achieved 93% (131/141) mass detection rate with 4.8 FPs/image (1,296/270). In the present study, the textural signature of the detected location is analyzed for possible false positive reduction. For texture analysis, HAAR wavelet and MPEG-7 HTD textural features were extracted. In addition, the contribution of directional neighborhood (DN) features was studied as well. The extracted features were combined with a back-propagation artificial neural network (BPANN) to discriminate true masses from false positives. Using a database of 1,427 suspicious seeds (131 true masses and 1,296 FPs) and a 5-fold cross-validation sampling scheme, the ROC area index of the BPNN using the different sets of features were as follows: A_z(HAAR)=0.87±0.01, A_z(HTD)=0.91±0.02, A_z(DN)=0.84±0.01. Averaging the scores of the three BPANNs resulted in statistically significantly better performance A_z(ALL)=0.94±0.01. At 95% sensitivity, the FP rate was reduced by 77.5%. The overall performance of the system after incorporation of textural and directional features was 87.9% sensitivity for malignant masses at 1.1 FPs/image.

An important purpose of a CAD system is that it can serve as a second reader to alert radiologists to subtle cancers that may be overlooked. In this study, we are developing new computer vision techniques to improve the detection performance for subtle masses on prior mammograms. A data set of 159 patients containing 318 current mammograms and 402 prior mammograms was collected. A new technique combining gradient field analysis with Hessian analysis was developed to prescreen for mass candidates. A suspicious structure in each identified location was initially segmented by seed-based region growing and then refined by using an active contour method. Morphological, gray level histogram and run-length statistics features were extracted. Rule-based and LDA classifiers were trained to differentiate masses from normal tissues. We randomly divided the data set into two independent sets; one set of 78 cases for training and the other set of 81 cases for testing. With our previous CAD system, the case-based sensitivities on prior mammograms were 63%, 48% and 32% at 2, 1 and 0.5 FPs/image, respectively. With the new CAD system, the case-based sensitivities were improved to 74%, 56% and 35%, respectively, at the same FP rates. The difference in the FROC curves was statistically significant (p<0.05 by AFROC analysis). The performances of the two systems for detection of masses on current mammograms were comparable. The results indicated that the new CAD system can improve the detection performance for subtle masses without a trade-off in detection of average masses.

This work utilizes a novel pipeline for the computer-aided detection (CAD) of colonic polyps, assisting radiologists in locating polyps when using a virtual colonoscopy system. Our CAD pipeline automatically detects polyps while reducing the number of false positives (FPs). It integrates volume rendering and conformal colon flattening with texture and shape analysis. The colon is first digitally cleansed, segmented, and extracted from the CT dataset of the abdomen. The colon surface is then mapped to a 2D rectangle using conformal mapping. Using this colon flattening method, the CAD problem is converted from 3D into 2D. The flattened image is rendered using a direct volume rendering of the 3D colon dataset with a translucent transfer function. Suspicious polyps are detected by applying a clustering method on the 2D volume rendered image. The FPs are reduced by analyzing shape and texture features of the suspicious areas detected by the clustering step. Compared with shape-based methods, ours is much faster and much more efficient as it avoids computing curvature and other shape parameters for the whole colon wall. We tested our method with 178 datasets and found it to be 100% sensitive to adenomatous polyps with a low rate of FPs. The CAD results are seamlessly integrated into a virtual colonoscopy system, providing the radiologists with visual cues and likelihood indicators of areas likely to contain polyps, and allowing them to quickly inspect the suspicious areas and further exploit the flattened colon view for easy navigation and bookmark placement.

We evaluate and improve an existing curvature-based region growing algorithm for colonic polyp detection for our CT colonography (CTC) computer-aided detection (CAD) system by using Pareto fronts. The performance of a polyp detection algorithm involves two conflicting objectives, minimizing both false negative (FN) and false positive (FP) detection rates. This problem does not produce a single optimal solution but a set of solutions known as a Pareto front. Any solution in a Pareto front can only outperform other solutions in one of the two competing objectives. Using evolutionary algorithms to find the Pareto fronts for multi-objective optimization problems has been common practice for years. However, they are rarely investigated in any CTC CAD system because the computation cost is inherently expensive. To circumvent this problem, we have developed a parallel program implemented on a Linux cluster environment. A data set of 56 CTC colon surfaces with 87 proven positive detections of polyps sized 4 to 60 mm is used to evaluate an existing one-step, and derive a new two-step region growing algorithm. We use a popular algorithm, the Strength Pareto Evolutionary Algorithm (SPEA2), to find the Pareto fronts. The performance differences are evaluated using a statistical approach. The new algorithm outperforms the old one in 81.6% of the sampled Pareto fronts from 20 simulations. When operated at a suitable sensitivity level such as 90.8% (79/87) or 88.5% (77/87), the FP rate is decreased by 24.4% or 45.8% respectively.

Colon cancer is a widespread disease and, according to the American Cancer Society, it is estimated that in 2006 more than 55,000 people will die of colon cancer in the US. However, early detection of colorectal polyps helps to drastically reduces mortality. Computer-Aided Detection (CAD) of colorectal polyps is a tool that could help physicians finding such lesions in CT scans of the colon. In this paper, we present the first phase, candidate generation (CG), of our technique for the detection of colonic polyp candidate locations in CT colonoscopy. Since polyps typically appear as protrusions on the surface of the colon, our cutting-plane algorithm identifies all those areas that can be "cut-off" using a plane. The key observation is that for any protruding lesion there is at least one plane that cuts a fragment off. Furthermore, the intersection between the plane and the polyp will typically be small and circular. On the other hand, a plane cannot cut a small circular cross-section from a wall or a fold, due to their concave or elongated paraboloid morphology, because these structures yield cross-sections that are much larger or non-circular. The algorithm has been incorporated as part of a prototype CAD system. An analysis on a test set of more than 400 patients yielded a high per-patient sensitivity of 95% and 90% in clean and tagged preparation respectively for polyps ranging from 6mm to 20mm in size.

Electronic cleansing (EC) is an emerging method for segmentation of fecal materials, which are tagged by an X-ray-opaque oral contrast agent in CT colonography (CTC) images, effectively removing them for digital cleansing of the colon. Due to the partial volume effect, voxels at the boundary between the lumen and tagged materials, called the L-T boundary, not only have CT values that are close to those of soft-tissue structures, but also have gradient values that are similar to a thin soft-tissue structure that is sandwiched between the tagged region and the lumen, which we call a tagging-tissue-air layer. Degradation of thin colonic wall and folds, as well as creation of pseudo soft-tissue structures at the periphery of tagged regions, are main artifacts in existing EC approaches, which tend to use a gradient-based method to delineate tagged regions. In this study, we developed a novel delineation method of tagged regions by applying local iso-surface roughness. The local iso-surface roughness is defined by the sum of differences between the local curvedness at adjacent scales over all scales. In our approach, the roughness values around the periphery of the tagged regions are integrated into the speed function of a level-set segmentation method for delineation of the tagged regions. As a result, L-T boundaries are subtracted along with tagged regions, whereas the thin soft-tissue structures within the tagging-tissue-air layers are preserved. Application of our computer-aided detection (CAD) scheme showed that the use of the new EC method substantially reduced the number of false-positive detections compared with that of our previous gradient-based method.

Fecal-tagging CT colonography (CTC) presents an opportunity to minimize colon cleansing while maintaining high diagnostic accuracy for the detection of colorectal lesions. However, the pseudo-enhancement introduced by tagging agents presents several problems for the application of computer-aided detection (CAD). We developed a correction method that minimizes pseudo-enhancement in CTC data by modeling of the pseudo-enhancement as a cumulative Gaussian energy distribution. The method was optimized by use of an anthropomorphic colon phantom, and its effect on our fully automated CAD scheme was tested by use of leave-one-patient-out evaluation on 23 clinical CTC cases with reduced colon cleansing based upon dietary fecal tagging. There were 28 colonoscopy-confirmed polyps ≥6 mm. Visual evaluation indicated that the method reduced CT attenuation of pseudo-enhanced polyps to standard soft-tissue Hounsfield unit (HU) range without affecting untagged regions. At a 90% detection sensitivity for polyps ≥6 mm, CAD yielded 8.5 false-positive (FP) detections and 3.9 FP detections per volumetric scan without and with the application of the pseudo-enhancement correction method. These results indicate that the pseudo-enhancement correction method is a potentially useful pre-processing step for automated detection of polyps in fecal-tagging CTC, and that CAD can yield a high detection sensitivity with a relatively low FP rate in CTC with patient-friendly reduced colon preparation.

Colon motility disorders are a very common problem. A precise diagnosis with current methods is almost unachievable. This makes it extremely difficult for the clinical experts to decide for the right intervention such as colon resection. The use of cine MRI for visualizing the colon motility is a very promising technique. In addition, if image segmentation and qualitative motion analysis provide the necessary tools, it could provide the appropriate diagnostic solution. In this work we defined necessary steps in the image processing workflow to gain valuable measurements for a computer aided diagnosis of colon motility disorders. For each step, we developed methods to deal with the dynamic image data. There is need for compensating the breathing motion since no respiratory gating could be used. We segment the colon using a graph cuts approach in 2D and 3D for further analysis and visualization. The analysis of the large bowel motility is done by tracking the extension of the colon during a propagating peristaltic wave. The main objective of this work is to extract a motion model to define a clinical index that can be used in diagnosis of large bowel motility dysfunction. We aim at the classification and localization of such pathologies.

CT colonography, a prevalent tool to diagnose colon cancer in its early stages, is often limited by bad distention, or retained fluids, which will cause segments of the colon to be impossible to process by CAD tools. By scanning patients in both prone and supine positions, collapsed segments and retained fluids will not be in the same place in both images, increasing the length of the colon that can be processed correctly. In order to fully use these two scans, they must be registered, so that a lesion identified on one of them can be mapped to the other, thus increasing sensitivity and specificity of CAD tools. The surface of the colon is however large (more than half a million vertices on our images), and has no canonical shape, which makes atlases and other widely used registration algorithms non optimal. We present in this paper a fast method to register the colon surface between prone and supine scans using landmarks present on the colon, the teniae coli. Our method is composed of three steps. First, we register the body, based on manually placed landmarks. Then we register the three teni&tildeae; coli, and, from this registration, we compute a deformation field for each vertex of the colon surface. We tested our method on 5 cases, by measuring the RMS error after body registration, quantifying the intrisic movement of the colon, and after colon surface registration. The RMS error was reduced from 1.8 cm to 0.49 cm, a reduction of 71%.

Normal function of the gastrointestinal tract involves the coordinated activity of several cell types Human disorders of motor function of the gastrointestinal tract are often associated with changes in the number of these cells. For example, in diabetic patients, abnormalities in gastrointestinal transit are associated with changes in nerves and interstitial cells of Cajal (ICC), two key cells that generate and regulate motility. ICC are cells of mesenchymal origin that function as pacemakers and amplify neuronal signals in the gastrointestinal tract. Quantifying the changes in number of specific cell types in tissues from patients with motility disorders is challenging and requires immunolabeling for specific antigens. The shape of nuclei differs between the cell types in the wall of the gastrointestinal tract. Therefore the objective of this study was to determine whether cell nuclei can be classified by analyzing the 3D morphology of the nuclei. Furthermore, the orientation of the long axis of nuclei changes within and between the muscle layers. These features can be used to classify and differentially label the nuclei in confocal volume images of the tissue by computing the principal axis of the coordinates of the set of voxels forming each nucleus and thereby to identify cells by their nuclear morphology. Using this approach, we were able to separate and quantify nuclei in the smooth muscle layers of the tissue. Therefore we conclude that computer-aided classification of cell nuclei can be used to identify changes in the cell types expressed in gastrointestinal smooth muscle.

Cervical Intraepithelial Neoplasia (CIN) is a precursor to invasive cervical cancer, which annually accounts for about 3700 deaths in the United States and about 274,000 worldwide. Early detection of CIN is important to reduce the fatalities due to cervical cancer. While the Pap smear is the most common screening procedure for CIN, it has been proven to have a low sensitivity, requiring multiple tests to confirm an abnormality and making its implementation impractical in resource-poor regions. Colposcopy and cervicography are two diagnostic procedures available to trained physicians for non-invasive detection of CIN. However, many regions suffer from lack of skilled personnel who can precisely diagnose the bio-markers due to CIN. Automatic detection of CIN deals with the precise, objective and non-invasive identification and isolation of these bio-markers, such as the Acetowhite (AW) region, mosaicism and punctations, due to CIN. In this paper, we study and compare three different approaches, based on Mathematical Morphology (MM), Deterministic Annealing (DA) and Gaussian Mixture Models (GMM), respectively, to segment the AW region of the cervix. The techniques are compared with respect to their complexity and execution times. The paper also presents an adaptive approach to detect and remove Specular Reflections (SR). Finally, algorithms based on MM and matched filtering are presented for the precise segmentation of mosaicism and punctations from AW regions containing the respective abnormalities.

Compared to histopathological methods cancer can be detected earlier, specimens can be obtained easier and with less discomfort for the patient by cytopathological methods. Their downside is the time needed by an expert to find and select the cells to be analyzed on a specimen. To increase the use of cytopathological diagnostics, the cytopathologist has to be supported in this task. DNA image cytometry (DNA-ICM) is one important cytopathological method that measures the DNA content of cells based on the absorption of light within Feulgen stained cells. The decision whether or not the patient has cancer is based on the histogram of the DNA values. To support the cytopathologist it is desirable to replace manual screening of the specimens by an automatic selection of relevant cells for DNA-ICM. This includes automated acquisition and segmentation of focused cells, a recognition of cell types, and a selection of cells to be measured. As a step towards automated cell type detection we show the discrimination of cell types in serous effusions on a selection of about 3, 100 manually classified cells. We present a set of 112 features and the results of feature selection with ranking and a floating-search method combined with different objective functions. The validation of the best feature sets with a k-nearest neighbor and a fuzzy k-nearest neighbor classifier on a disjoint set of cells resulted in classification rates of 96% for lymphocytes and 96.8% for the diagnostically relevant cells (mesothelial+ cells), which includes benign and malign mesothelial cells and metastatic cancer cells.

Most methods for classifier design assume that the training samples are drawn independently and identically from an unknown data generating distribution (i.i.d.), although this assumption is violated in several real life problems. Relaxing this i.i.d. assumption, we develop training algorithms for the more realistic situation where batches or sub-groups of training samples may have internal correlations, although the samples from different batches may be considered to be uncorrelated; we also consider the extension to cases with hierarchical--i.e. higher order--correlation structure between batches of training samples. After describing efficient algorithms that scale well to large datasets, we provide some theoretical analysis to establish their validity. Experimental results from real-life Computer Aided Detection (CAD) problems indicate that relaxing the i.i.d. assumption leads to statistically significant improvements in the accuracy of the learned classifier.

Our purpose was to reduce false-positive (FP) detections generated by a computerized lesion detection scheme by using a "mixture of expert" massive-training artificial neural networks (MTANNs). Multiple MTANNs were trained with "ensemble training" for reduction of diverse types of non-lesions. We started from a seed MTANN trained with lesions and non-lesions of a seed type. We applied the trained seed MTANN to lesions and various types of non-lesions to analyze the weakness of the seed MTANN. We arranged the output scores of the MTANN for lesions and non-lesions in an ascending order to form the score scale representing the degree of difficulty in distinction between lesions and non-lesions by the seed MTANN. The score scale was divided into several segments, and ten non-lesions were sampled from the center of each segment so that sets of non-lesion samples covered diverse difficulties. We trained several MTANNs with several sets of non-lesions so that each MTANN became an expert for the non-lesions at a certain level of difficulty. We then combined expert MTANNs with a mixing ANN to form a "mixture of expert" MTANNs. Our database consisted of CT colonography datasets acquired from 100 patients, including 26 polyps. We applied our initial CAD scheme to this CTC database. FP sources included haustral folds, stool, colonic walls, the ileocecal valves, and rectal tubes. The mixture of expert MTANNs distinguished all polyps correctly from more than 50% of the non-polyps. Thus, the mixture of expert MTANNs was able reduce one half of the FPs generated by a computerized polyp detection scheme while the original sensitivity was maintained. We compare the effectiveness of ensemble training with that of training with manually selected cases. The performance of the MTANNs with ensemble training was superior to that of the MTANNs trained with manually selected cases.

Size is an important metric for pulmonary nodule characterization. Furthermore, it is an important parameter in measuring the performance of computer aided detection systems since they are always qualified with respect to a given size range of nodules. The first 120 whole-lung CT scans documented by the Lung Image Database Consortium using their protocol for nodule evaluation were used in this study. For documentation, each inspected lesion was reviewed independently by four expert radiologists and, when a lesion was considered to be a nodule larger than 3mm, the radiologist provided boundary markings in each image in which the nodule was contained. Three size metrics were considered: a uni-dimensional and a bi-dimensional measure on a single image slice and a volumetric measurement based on all the image slices. In this study we analyzed the boundary markings of these nodules in the context of these three size metrics to characterize the inter-radiologist variation and to examine the difference between these metrics. A data set of 63 nodules each having four observations was analyzed for inter-observer variation and an extended set of 252 nodules each having at least one observation was analyzed for the difference between the metrics. A very high inter-observer variation was observed for all these metrics and also a very large difference among the metrics was observed.

The LIDC is developing a publicly available database of thoracic computed tomography (CT) scans as a medical imaging research resource. A unique multi-center data collection process and communication system were developed to share image data and to capture the location and spatial extent of lung nodules as marked by expert radiologists. A two-phase data collection process was designed to allow multiple radiologists at different centers to asynchronously review and annotate each CT image series. Four radiologists reviewed each case using this process. In the first or "blinded" phase, each radiologist reviewed the CT series independently. In the second or "unblinded" review phase, the results from all four blinded reviews are compiled and presented to each radiologist for a second review. This allows each radiologist to review their own annotations along with those of the other radiologists. The results from each radiologist's unblinded review were compiled to form the final unblinded review. There is no forced consensus in this process. An XML-based message system was developed to communicate the results of each reading. This two-phase data collection process was designed, tested and implemented across the LIDC. It has been used for more than 130 CT cases that have been read and annotated by four expert readers and are publicly available at (http://ncia.nci.nih.gov). A data collection process was developed, tested and implemented that allowed multiple readers to review each case multiple times and that allowed each reader to observe the annotations of other readers.

In computerized nodule detection systems on CT scans, many features that are useful for classifying whether a nodule candidate identified by prescreening is a true positive depend on the shape of the segmented object. We designed two segmentation algorithms for detailed delineation of the boundaries for nodule candidates. The first segmentation technique was a three-dimensional (3D) region-growing (RG) method which grew the object across multiple CT sections. The second technique was based on a 3D active contour (AC) model. A training set of 94 CT scans was used for algorithm design. An independent set of 62 scans, each read by multiple radiologists, was used for testing. Thirty-three scans were collected from patient files at the University of Michigan and 29 scans by the Lung Imaging Database Consortium (LIDC). In this study, we concentrated on the detection of internal lung nodules having a size ≥3 mm that were not pure ground-glass opacities. Of the lesions marked by one or multiple radiologists, 124 nodules satisfied these criteria and were considered true nodules. The performance of the detection system in the AC feature space, RG feature space, and the combined feature space were compared using free-response receiver operating curves (FROC). The FROC curve using the combined feature space was significantly higher than that using the RG feature space or the AC feature space alone (p=0.02 and 0.03, respectively). At a sensitivity of 70% for internal non-GGO nodules, the FP rates were 2.2, 2.2, and 1.5 per scan, respectively, for the RG, AC, and the combined methods. Our results indicate that the 3D AC algorithm can provide useful features to improve nodule detection on CT scans.

Neural networks (NN) are typically developed to minimize the squared difference between the network's output and the target value for a set of training patterns; namely the mean squared error (MSE). However, lower MSE does not necessarily translate into a clinically more useful decision model. The purpose of this study was to investigate the particle swarm optimization (PSO) algorithm as an alternative way of NN optimization with clinically relevant objective functions (e.g., ROC and partial ROC area indices). The PSO algorithm was evaluated with respect to a NN-based CAD system developed to discriminate mammographic regions of interest (ROIs) that contained masses from normal regions based on 8 computer-extracted morphology-oriented features. Neural networks were represented as points (particle locations) in a D-dimensional search/optimization space where each dimension corresponded to one adaptable NN parameter. The study database of 1,337 ROIs (681 with masses, 656 normal) was split into two subsets to implement two-fold cross-validation sampling scheme. Neural networks were optimized with the PSO algorithm and the following objective functions (1) MSE, (2) ROC area index AUC, and (3) partial ROC area indices _TPFAUC with TPF=0.90 and TPF=0.98. For comparison, performance of neural networks of the same architecture trained with the traditional backpropagation algorithm was also evaluated. Overall, the study showed that when the PSO algorithm optimized network parameters using a particular training objective, the NN test performance was superior with respect to the corresponding performance index. This was particularly true for the partial ROC area indices where statistically significant improvements were observed.

Reliable and accurate methods for objective quantitative assessment of parenchymal alterations in the lung are necessary for diagnosis, treatment and follow-up of pulmonary diseases. Two major types of alterations are pulmonary emphysema and fibrosis, emphysema being characterized by abnormal enlargement of the air spaces distal to the terminal, nonrespiratory bronchiole, accompanied by destructive changes of the alveolar walls. The main characteristic of fibrosis is coursening of the interstitial fibers and compaction of the pulmonary tissue. With the ability to display anatomy free from superimposing structures and greater visual clarity, Multi-Detector-CT has shown to be more sensitive than the chest radiograph in identifying alterations of lung parenchyma. In automated evaluation of pulmonary CT-scans, quantitative image processing techniques are applied for objective evaluation of the data. A number of methods have been proposed in the past, most of which utilize simple densitometric tissue features based on the mean X-ray attenuation coefficients expressed in terms of Hounsfield Units [HU]. Due to partial volume effects, most of the density-based methodologies tend to fail, namely in cases, where emphysema and fibrosis occur within narrow spatial limits. In this study, we propose a methodology based upon the topological assessment of graylevel distribution in the 3D image data of lung tissue which provides a way of improving quantitative CT evaluation. Results are compared to the more established density-based methods.

Computer aided quantification of emphysema in high resolution CT data is based on identifying low attenuation areas below clinically determined Hounsfield thresholds. However, the emphysema quantification is prone to error since a gravity effect can influence the mean attenuation of healthy lung parenchyma up to ± 50 HU between ventral and dorsal lung areas. Comparing ultra-low-dose (7 mAs) and standard-dose (70 mAs) CT scans of each patient we show that measurement of the ventrodorsal gravity effect is patient specific but reproducible. It can be measured and corrected in an unsupervised way using robust fitting of a linear function.

We describe a system for the automatic detection of rib metastasis in thoracic CT volume. Rib metastasis manifest itself as alterations of bone intensities or shapes, and the detection of these alterations is the goal of the algorithm. According to the tubular shape of the rib structures, the detection is based on the construction of 2D cross-sections planes along the full lengths of each of the individual ribs. The set of planes is orthogonal to the rib centerline, with is extracted by a previously developed segmentation algorithm based on recursive tracing. On each of these planes, a 2D image is constructed by interpolation in the region of interest around the centerline intersection and the plane. From this image the cortical and trabecular bones are segmented separately. The appearance and geometric properties of the bone structures are analyzed and categorized according to a set of rules that summarize the possible variation types due to metastasis. The features extracted from the cross-sections along a short length of the centerline are jointly evaluated. A positive detection is accepted only if the alteration of shape and appearance is consistent with a number of consecutive cross-sections along the rib centerline.

Automated methods of detecting lung disease typically involve the following: 1) Subdividing the lung into small regions of interest (ROIs). 2) Calculating the features of these small ROIs. 3) Applying a machine learnt classifier to determine the class of each ROI. When the number of features that need to be calculated is large, as in the case of filter bank methods or in methods calculating a large range of textural properties, the classification can run quite slowly. This is even more noticeable when a number of disease patterns are considered. In this paper, we investigate the possibility of using a cascade of classifiers to concentrate the processing power on promising regions. In particular, we focused on the detection of the honeycombing disease pattern. We used knowledge of the appearance and the distribution of honeycombing to selectively classify ROIs. This avoids the need to explicitly classify all ROIs in the lung; making the detection process more effcient. We evaluated the performance of the system over 42 HRCT slices from 8 different patients and show that the system performs the task of detecting honeycombing with a high degree of accuracy (accuracy = 86.2%, sensitivity = 90.0%, specificity = 82.2%).

The goal of this study was to develop an automated method to identify corresponding nodules in serial CT scans for interval change analysis. The method uses the rib centerlines as the reference for initial nodule registration. From an automatically-identified starting point near the spine, each rib is locally tracked and segmented by expectation-maximization. The ribs are automatically labeled, and the centerlines are estimated using skeletonization. 3D rigid affine transformation is used to register the individual ribs in the reference and target scans. For a given nodule in the reference scan, a search volume of interest (VOI) in the target scan is defined by using the registered ribs. Template matching guided by the normalized cross-correlation between the nodule template and target locations within the search VOI is used for refining the registration. The method was evaluated on 48 CT scans from 20 patients. The slice thickness ranged from 0.625 to 7 mm, and the in-plane pixel size from 0.556 to 0.82 mm. Experienced radiologists identified 101 pairs of nodules. Two metrics were used for performance evaluation: 1) the Euclidean distance between the nodule centers identified by the radiologist and the computer registration, and 2) a volume overlap measure defined as the intersection of the VOIs identified by the radiologist and the computer registration relative to the radiologist's VOI. The average Euclidean distance error was 2.7 ± 3.3 mm. Only 2 pairs had an error >10 mm. The average volume overlap measure was 0.71 ± 0.24. Eight-three out of 101 pairs had overlap ratios > 0.5 and only 2 pairs had no overlap.

We study the problem of classifying brain tumors as benign or malignant using information from magnetic resonance (MR) imaging and magnetic resonance spectroscopy (MRS) to assist in clinical diagnosis. The proposed approach consists of several steps including segmentation, feature extraction, feature selection, and classification model construction. Using an automated segmentation technique based on fuzzy connectedness we accurately outline the tumor mass boundaries in the MR images so that further analysis concentrates on these regions of interest (ROIs). We then apply a concentric circle technique on the ROIs to extract features that are utilized by the classification algorithms. To remove redundant features, we perform feature selection where only those features with discriminatory information (among classes) are used in the model building process. The involvement of MRS features further improves the classification accuracy of the model. Experimental results demonstrate the effectiveness of the proposed approach in classifying brain tumors in MR images.

Dynamic contrast enhanced breast MRI (DCE BMRI) is an emerging tool for breast cancer diagnosis. There is a clear clinical demand for computer-aided diagnosis (CADx) tools to support radiologists in the diagnostic reading process of DCE BMRI studies. A crucial step in a CADx system is the segmentation of tumors, which allows for accurate assessment of the 3D lesion size and morphology. In this paper we propose a semiautomatic segmentation procedure for suspicious breast lesions. The proposed methodology consists of four steps: (1) Robust seed point selection. This interaction mode ensures robustness of the segmentation result against variations in seed-point placement. (2) Automatic intensity threshold estimation in the subtraction image. (3)Connected component analysis based on the estimated threshold. (4) A post-processing step that includes non-enhancing portions of the lesion into the segmented area and removes attached vessels. The proposed methodology was applied to DCE BMRI data acquired at different institutions using different protocols.

In this study, we investigated the effect of different patient calibration methods on the performance of our CAD system when discriminating prostate cancer from non-malignant suspicious enhancing areas in the peripheral zone and the normal peripheral zone. Our database consisted of 34 consecutive patients with histologically proven adenocarcinoma of the prostate. Both carcinoma and normal tissue were annotated on MR images by a radiologist and a researcher using whole mount step-section histopathology as standard of reference. The annotated regions were used as regions of interest in the contrast enhanced MRI images. A feature set comprising pharmacokinetic parametes was extracted from the ROIs to train a support vector machine as classifier. The output of the classifier was used as a measure of likelihood of malignancy. General performance of the scheme was evaluated using the area under the ROC curve. The diagnostic accuracy obtained for differentiating normal peripheral zone and non-malignant suspicious enhancing areas from malignant lesions was 0.88 (0.81-0.95) when per patient calibration was performed, whereas fixed calibration resulted in a diagnostic accuracy of 0.77 (0.69-0.85). These preliminary results indicate that when per patient calibration is used, the performance is improved with statistical significance (p=0.026).

Brain tumor segmentation and quantification from MR images is a challenging task. The boundary of a tumor and its volume are important parameters that can have direct impact on surgical treatment, radiation therapy, or on quantitative measurements of tumor regression rates. Although a wide range of different methods has already been proposed, a commonly accepted approach is not yet established. Today, the gold standard at many institutions still consists of a manual tumor outlining, which is potentially subjective, and a time consuming and tedious process. We propose a new method that allows for fast multispectral segmentation of brain tumors. An efficient initialization of the segmentation is obtained using a novel probabilistic intensity model, followed by an iterative refinement of the initial segmentation. A progressive region growing that combines probability and distance information provides a new, flexible tumor segmentation. In order to derive a robust model for brain tumors that can be easily applied to a new dataset, we retain information not on the anatomical, but on the global cross-subject intensity variability. Therefore, a set of multispectral histograms from different patient datasets is registered onto a reference histogram using global affine and non-rigid registration methods. The probability model is then generated from manual expert segmentations that are transferred to the histogram feature domain. A forward and backward transformation of a manual segmentation between histogram and image domain allows for a statistical analysis of the accuracy and robustness of the selected features. Experiments are carried out on patient datasets with different tumor shapes, sizes, locations, and internal texture.

A system for automatic detection of pelvic lymph nodes is developed by incorporating complementary information extracted from multiple MR sequences. A single MR sequence lacks sufficient diagnostic information for lymph node localization and staging. Correct diagnosis often requires input from multiple complementary sequences which makes manual detection of lymph nodes very labor intensive. Small lymph nodes are often missed even by highly-trained radiologists. The proposed system is aimed at assisting radiologists in finding lymph nodes faster and more accurately. To the best of our knowledge, this is the first such system reported in the literature. A 3-dimensional (3D) MR angiography (MRA) image is employed for extracting blood vessels that serve as a guide in searching for pelvic lymph nodes. Segmentation, shape and location analysis of potential lymph nodes are then performed using a high resolution 3D T1-weighted VIBE (T1-vibe) MR sequence acquired by Siemens 3T scanner. An optional contrast-agent enhanced MR image, such as post ferumoxtran-10 T2*-weighted MEDIC sequence, can also be incorporated to further improve detection accuracy of malignant nodes. The system outputs a list of potential lymph node locations that are overlaid onto the corresponding MR sequences and presents them to users with associated confidence levels as well as their sizes and lengths in each axis. Preliminary studies demonstrates the feasibility of automatic lymph node detection and scenarios in which this system may be used to assist radiologists in diagnosis and reporting.

Background: Reported error rates for initial clinical diagnosis in parkinsonian disorders can reach up to 35%. Reducing this initial error rate is an important research goal. The objective of this work is to evaluate the ability of an automated MR-based classification technique in the differential diagnosis of Parkinson's disease (PD), multiple systems atrophy (MSA) and progressive supranuclear palsy (PSP). Methods: A total of 172 subjects were included in this study: 152 healthy subjects, 10 probable PD patients and 10 age-matched patients with diagnostic of either probable MSA or PSP. T1-weighted (T1w) MR images were acquired and subsequently corrected, scaled, resampled and aligned within a common referential space. Tissue transformation and deformation features were then automatically extracted. Classification of patients was performed using forward, stepwise linear discriminant analysis within a multidimensional transformation/deformation feature space built from healthy subjects data. Leave-one-out classification was used to avoid over-determination. Findings: There were no age difference between groups. Highest accuracy (agreement with long-term clinical follow-up) of 85% was achieved using a single MR-based deformation feature. Interpretation: These preliminary results demonstrate that a classification approach based on quantitative parameters of 3D brainstem morphology extracted automatically from T1w MRI has the potential to perform differential diagnosis of PD versus MSA/PSP with high accuracy.

There is widespread clinical interest in the study of pulmonary nodules for early diagnosis of lung cancer. These nodules can be broadly classified into one of three types, solid, nonsolid and part-solid. Solid nodules have been extensively studied, while little research has focused on the characterization of nonsolid and part-solid nodules. Nonsolid nodules have an appearance in high-resolution CT consisting of voxels only slightly more dense than that of the surrounding lung parenchyma. For the solid nodule, robust techniques are available to estimate growth rate and this is commonly used to distinguish benign from malignant. For the nonsolid types, these techniques are less well developed. In this research, we propose an automated volumetric segmentation method for nonsolid nodules that accurately determines a nonsolid nodule's growth rate. Our method starts with an initial noise-filtering stage in the parenchyma region. Each voxel is then classified into one of three tissue types; lung parenchyma, nonsolid and solid. Removal of vessel attachments to the lesion is achieved with the use of a filter that focuses on vessel characteristics. Our results indicate that the automated method is more consistent than the radiologist with a median growth consistency of 1.87 compared to 3.12 for the radiologist on a database of 25 cases.

A new lung nodule simulation model was designed to create and insert synthetic solid lung nodules, with shapes and density similar to real nodules, into normal MDCT chest exams. Nodule shapes were modeled using linearly deformed superquadrics with added randomly generated high dimensional deformations. Nodule density statistics and attenuation profiles were extracted from a group of real nodule samples, by dissecting each real nodule digitally layer by layer from the border to the core. A nodule created with modeled shape and density was inserted into real CT images by creating volume average layers using weighted averaging between nodule density and background density for each voxel. The nodule simulation model was validated both subjectively by human experts and quantitatively by comparing density attenuation profiles of simulated nodules with real nodules. These validation studies demonstrated a high level of similarity between the synthetic nodules and real nodules. This nodule simulation model was used to create objective test databases for use in evaluating a CAD system. The evaluation study showed that the CAD system was accurate in detection and volume measurement for isolated nodules, and also performed relatively well for juxta-vascular nodules. The CAD system also demonstrated stable performances for different dosages.

The automated detection of lung nodules in CT scans is an important problem in computer-aided diagnosis. In this paper an approach to nodule candidate detection is presented which utilises the local image features of shape index and curvedness. False-positive candidates are removed by means of a two-step approach using kNN classification. The kNN classifiers are trained using features of the image intensity gradients and grey-values in addition to further measures of shape index and curvedness profiles in the candidate regions. The training set consisted of data from 698 scans while the independent test set comprised a further 142 images. At 84% sensitivity an average of 8.2 false-positive detections per scan were observed.

A CAD system was developed to extract and analyze features from corresponding malignant and benign lung nodules on temporal pairs of CT scans. The lung nodules on the current and prior CT scans were automatically segmented using a 3-dimensional (3D) active contour model. Three-dimensional run length statistics (RLS) texture features, 3D morphological and gray-level features were extracted from each nodule. In addition, 3D nodule profile features (PROF) that describe the gray level variation inside and outside the nodule surface were extracted by estimating the gradient magnitude values along the radial vectors from the nodule centroid to a band of voxels surrounding the nodule surface. Interval change features were calculated as the difference between the corresponding features extracted from the prior and the current scans of the same nodule. Stepwise feature selection with simplex optimization was used to select the best feature subset from the feature space that combined both the interval change features and features from the single current exam. A linear discriminant classifier was used to merge the selected features for classification of malignant and benign nodules. In this preliminary study, a data set of 103 nodule temporal pairs (39 malignant and 64 benign) was used. A leave-one-case-out resampling scheme was used for feature selection and classification. An average of 5 features was selected from the training subsets. The most frequently selected features included a difference PROF feature and 4 RLS features. The classifier achieved a test Az of 0.85±0.04. In comparison a classifier using features extracted from the current CT scans alone achieved a test Az of 0.78±0.05. This study indicates that our CAD system using interval change information is useful for classification of lung nodules on CT scans.

Recent advance of imaging technology has brought new challenges and opportunities for automatic and quantitative analysis of medical images. With broader accessibility of more imaging modalities for more patients, fusion of modalities/scans from one time point and longitudinal analysis of changes across time points have become the two most critical differentiators to support more informed, more reliable and more reproducible diagnosis and therapy decisions. Unfortunately, scan fusion and longitudinal analysis are both inherently plagued with increased levels of statistical errors. A lack of comprehensive analysis by imaging scientists and a lack of full awareness by physicians pose potential risks in clinical practice. In this paper, we discuss several key error factors affecting imaging quantification, studying their interactions, and introducing a simulation strategy to establish general error bounds for change quantification across time. We quantitatively show that image resolution, voxel anisotropy, lesion size, eccentricity, and orientation are all contributing factors to quantification error; and there is an intricate relationship between voxel anisotropy and lesion shape in affecting quantification error. Specifically, when two or more scans are to be fused at feature level, optimal linear fusion analysis reveals that scans with voxel anisotropy aligned with lesion elongation should receive a higher weight than other scans. As a result of such optimal linear fusion, we will achieve a lower variance than naïve averaging. Simulated experiments are used to validate theoretical predictions. Future work based on the proposed simulation methods may lead to general guidelines and error lower bounds for quantitative image analysis and change detection.

The purpose of this paper was to investigate the effects of integrating nodule 3D morphological features, texture features and functional dynamic contrast-enhanced features in differentiating between benign and malignant solitary pulmonary nodules (SPNs). In this study, 42 cases with solitary lung nodules were examined in this study. The dynamic CT helical scans were acquired image at five time intervals: prior to contrast injection (baseline) and then at 45, 90, 180, 300 seconds after administrating the contrast agent. The nodule boundaries were contoured by radiologists on all series. Using these boundaries, several types of nodule features were computed, including: 3D morphology and Shape Index of the nodule contrast intensity surface; Dynamic contrast related features; 3D texture features. AdaBoost was performed to select the best features. Logistic Regression Analysis (LRA) and AdaBoost were used to analyze the diagnostic accuracy of features in each feature category. The performance when integrating all feature types was also evaluated. For 42 patients, when using only six SI and 3D structural features, the accuracy of AdaBoost was 81.4%, with accuracies of AdaBoost using functional contrast related features (include 8 features) and texture features(include 18 features) were 65.1% and 69.1% respectively. After combining all types' features together, the overall accuracy was improved to over 88%. In conclusion: Combining 3D structural, textural and functional contrast features can provide a more comprehensive examination of the SPNs by coupling dynamic CT scan techniques with image processing to quantify multiple properties that relate to tumor geometry and tumor angiogenesis. This integration may assist radiologists in characterizing SPNs more accurately.

The purpose of this study was to investigate feasibility of computer-aided detection of masses and calcification clusters in breast tomosynthesis images and obtain reliable estimates of sensitivity and false positive rate on an independent test set. Automatic mass and calcification detection algorithms developed for film and digital mammography images were applied without any adaptation or retraining to tomosynthesis projection images. Test set contained 36 patients including 16 patients with 20 known malignant lesions, 4 of which were missed by the radiologists in conventional mammography images and found only in retrospect in tomosynthesis. Median filter was applied to tomosynthesis projection images. Detection algorithm yielded 80% sensitivity and 5.3 false positives per breast for calcification and mass detection algorithms combined. Out of 4 masses missed by radiologists in conventional mammography images, 2 were found by the mass detection algorithm in tomosynthesis images.

The purpose of this project is to study Computer Aided Detection (CADe) of breast masses for digital tomosynthesis. It is believed that tomosynthesis will show improvement over conventional mammography in detection and characterization of breast masses by removing overlapping dense fibroglandular tissue. This study used the 60 human subject cases collected as part of on-going clinical trials at Duke University. Raw projections images were used to identify suspicious regions in the algorithm's high-sensitivity, low-specificity stage using a Difference of Gaussian (DoG) filter. The filtered images were thresholded to yield initial CADe hits that were then shifted and added to yield a 3D distribution of suspicious regions. These were further summed in the depth direction to yield a flattened probability map of suspicious hits for ease of scoring. To reduce false positives, we developed an algorithm based on information theory where similarity metrics were calculated using knowledge databases consisting of tomosynthesis regions of interest (ROIs) obtained from projection images. We evaluated 5 similarity metrics to test the false positive reduction performance of our algorithm, specifically joint entropy, mutual information, Jensen difference divergence, symmetric Kullback-Liebler divergence, and conditional entropy. The best performance was achieved using the joint entropy similarity metric, resulting in ROC A_z of 0.87 ± 0.01. As a whole, the CADe system can detect breast masses in this data set with 79% sensitivity and 6.8 false positives per scan. In comparison, the original radiologists performed with only 65% sensitivity when using mammography alone, and 91% sensitivity when using tomosynthesis alone.

We are developing a computer-aided detection (CAD) system for masses on digital breast tomosynthesis mammograms (DBTs). The CAD system includes two parallel processes. In the first process, mass detection and feature analysis are performed in the reconstructed 3D DBT volume. A mass likelihood score is estimated for each mass candidate using a linear discriminant (LDA) classifier. In the second process, mass detection and feature analysis are applied to the individual projection view (PV) images. A mass likelihood score is estimated for each mass candidate using another LDA classifier. The mass likelihood images derived from the PVs are back-projected to the breast volume to estimate the 3D spatial distribution of the mass likelihood scores. The mass likelihood scores estimated by the two processes at the corresponding 3D location are then merged and evaluated using FROC analysis. In this preliminary study, a data set of 52 DBT cases acquired with a GE prototype system at the Massachusetts General Hospital was used. The LDA classifiers with stepwise feature selection were designed with leave-one-case-out resampling. In an FROC analysis, the CAD system for detection in the DBT volume alone achieved test sensitivities of 80% and 90% at an average FP rate of 1.6 and 3.0 per breast, respectively. In comparison, the average FP rates of the combined system were 1.2 and 2.3 per breast, respectively, at the same sensitivities. The combined system is a promising approach to improving mass detection on DBTs.

We have analyzed breast parenchymal texture in tomosynthesis images. Tomosynthesis is a novel x-ray imaging modality in which 3D images of the breast are reconstructed from multiple 2D x-ray source projection images acquired by varying the angle of the x-ray tube. Our ultimate goal is to examine the correlation between tomosynthesis texture descriptors and breast cancer risk. As a first step, we investigated the effect of tomosynthesis acquisition parameters on texture in the source projection images; this avoids the influence of the reconstruction algorithm. We computed statistical texture descriptors which have been shown in the literature to be highly indicative of breast cancer risk. We compared skewness, coarseness, and contrast computed from the central source projection images and the corresponding mammograms. Our analysis showed that differences exist between mammographic and tomosynthetic texture in projection images. Retroareolar ROIs in tomosynthesis images appeared to be less skewed with lower coarseness and higher contrast measures compared to mammograms; however, corresponding texture descriptors for tomosynthesis and mammography are correlated. Examination of the ROIs demonstrates that the texture in tomosynthesis source projections visually differs from the x-ray mammograms. We attribute this observation to acquisition differences, including radiation dose, compression force, and x-ray scatter. As with mammography, tomosynthesis parenchymal texture is related to the Gail-model cancer risk. Although preliminary, we believe that texture analysis of 3D breast tomosynthesis images will ultimately yield more accurate and precise measures of risk.

We present an automated method for delineation of coronary arteries from Cardiac CT Angiography (CTA) images. Coronary arteries are narrow blood vessels and when imaged using CTA, appear as thin cylindrical structures of varying curvature. This appearance is often affected by heart motion and image reconstruction artifacts. Moreover, when an artery is diseased, it may appear as a non-continuous structure of widely varying width and image intensity. Defining the boundaries of the coronary arteries is an important and necessary step for further analysis and diagnosis of coronary disease. For this purpose, we developed a method using cylindrical structure modeling. For each vessel segment a best fitting cylindrical template is found. By applying this technique sequentially along the vessel, its entire volume can be reconstructed. The algorithm is seeded with a manually specified starting point at the most distal discernible portion of an artery and then it proceeds iteratively toward the aorta. The algorithm makes necessary corrections to account for CTA image artifacts and is able to perform in diseased arteries. It stops when it identifies the vessels junction with the aorta. Five cardiac 3D CT angiography studies were used for algorithm validation. For each study, the four longest visually discernible branches of the major coronary arteries were evaluated. Central axes obtained from our automated method were compared with ground truth markings made by an experienced radiologist. In 75% of the cases, our algorithm was able to extract the entire length of the artery from single initialization.

A multistage algorithm is presented, whose components are based upon maximum likelihood estimation (MLE). From 3D scanning laser ophthalmoscope (SLO) image data, the algorithm finds the positions of the two anatomical boundaries of the eye's fundus that define the retina, which are the internal limiting membrane (ILM) and the retinal pigment epithelium (RPE). he retinal thickness is then calculated by subtraction. Retinal thickness is useful for indicating, assessing risk of, and following several diseases, including various forms of macular edema and cysts.

To facilitate the clinical diagnosis, surgical planning and after operation follow-up, a computer aided septal defect diagnosis and detection framework is proposed. The framework is consisted of four steps: image registration, flow balance measurement, heart wall tracking and septal defects detection. First, a global smooth constrained localized registration method is employed to register the image; Then, flow balance measurement is employed to determine the unbalance of in and out blood flow, which usually indicate the septal defects in the heart; After that, the wall of the heart is tracked using the same framework used in the registration to improve the efficiency and accuracy; Defects along septal are detected using a Bayesian based information fusion to analyze the profile lines from registered image, difference image and original grey image for the whole sequence (3D+T). The proposed method is tested using gated cardiac MRI, which is a well-established clinical diagnosis method for septal defect detection. Experimental results show that the proposed framework is able to successfully detect the septal defects and provide the visual assistance to the radiologist for further diagnosis. The proposed detection can be widely used in both clinical practice, surgical planning and after operation follow-up. To the best of our knowledge, the work is first such an effort.

The accurate interpretation of cardiac CT images is commonly hindered by the presence of motion artifacts. Since motion artifacts commonly can obscure the presence of coronary lesions, physicians must spend much effort analyzing images at multiple cardiac phases in order to determine which coronary structures are assessable for potential lesions. In this study, an artificial neural network (ANN) classifier was designed to assign assessability indices to calcified plaques in individual region-of-interest (ROI) images reconstructed at multiple cardiac phases from two cardiac scans obtained at heart rates of 66 bpm and 90 bpm. Six individual features (volume, circularity, mean intensity, margin gradient, velocity, and acceleration) were used for analyzing images. Visually-assigned assessability indices were used as a continuous truth, and jack-knife analysis with four testing sets was used to evaluate the performance of the ANN classifier. In a study in which all six features were inputted into the ANN classifier, correlation coefficients of 0.962 ± 0.006 and 0.935 ± 0.023 between true and ANN-assigned assessability indices were obtained for databases corresponding to 66 bpm and 90 bpm, respectively.

This paper presents a method for extracting multi-organs from four-phase contrasted CT images taken at different contrast timings (non-contrast, early, portal, and late phases). First, we apply a median filter to each CT image and align four-phase CT images by performing non-rigid volumetric image registration. Then, a three-dimensional joint histogram of CT values is computed from three-phase (early-, portal-, and late-) CT images. We assume that this histogram is a mixture of normal distributions corresponding to the liver, spleen, kidney, vein, artery, muscle, and bone regions. The EM algorithm is employed to estimate each normal distribution. Organ labels are assigned to each voxel using the mahalanobis distance measure. Connected component analysis is applied to correct the shape of each organ region. After that, the pancreas region is extracted from non-contrasted CT images in which other extracted organs and vessel regions are excluded. The EM algorithm is also employed for estimating the distribution of CT values inside the pancreas. We applied this method to seven cases of four-phase CT images. Extraction results show that the proposed method extracted multi-organs satisfactorily.

In the US, colorectal cancer is the second leading cause of all cancer deaths behind lung cancer. Colorectal polyps are the precursor lesions of colorectal cancer. Therefore, early detection of polyps and at the same time removal of these precancerous lesions is one of the most important goals of colonoscopy. To objectively document detection and removal of colorectal polyps for quality purposes, and to facilitate real-time detection of polyps in the future, we have initiated a computer-based research program that analyzes video files created during colonoscopy. For computer-based detection of polyps, texture based techniques have been proposed. A major limitation of the existing texture-based analytical methods is that they depend on a fixed-size analytical window. Such a fixed-sized window may work for still images, but is not efficient for analysis of colonoscopy video files, where a single polyp can have different relative sizes and color features, depending on the viewing position and distance of the camera. In addition, the existing methods do not consider shape features. To overcome these problems, we here propose a novel polyp region segmentation method primarily based on the elliptical shape that nearly all small polyps and many larger polyps possess. Experimental results indicate that our proposed polyp detection method achieves a sensitivity and specificity of 93% and 98%, respectively.

With the emergence of recent technology in breast ultrasound, sonographic image quality has changed profoundly. Most notably, the technique of real-time spatial compounding impacts the appearance of lesions and parenchyma. During image acquisition, spatial compounding can be turned on or off at the discretion of the radiologist, but this information is not stored along with the image data. The ability to distinguishing between lesions imaged with and without spatial compounding, using either single image features or a Bayesian neural net (BNN), was assessed using ROC analysis. Our database consisted of consecutively collected HDI5000 images of 129 lesions imaged without spatial compounding (357 images, cancer prevalence of 18%) and 370 lesions imaged with spatial compounding (965 images, cancer prevalence 15%). These were used in automated feature selection and BNN training. An additional 33 lesions were imaged for which identical views with and without spatial compounding were available (70 images, cancer prevalence 15%). These served as an independent test dataset. Lesions were outlined by a radiologist and image features, mathematically describing lesion characteristics, were calculated. In feature selection, the 4 best performing features were related to gradient strength and entropy. The average gradient strength within a lesion obtained an area under the ROC curve (AUC) of 0.78 in the task of distinguishing lesions imaged with and without spatial compounding. The BNN, using 4 features, achieved an AUC on the independent test dataset of 0.98 in this task. The sonographic appearance of breast lesions is affected by spatial compound imaging and lesion features may be used to automatically separate images as obtained with or without this technique. In computer-aided diagnosis (CADx), it will likely be beneficial to separate images as such before using separate classifiers for assessment of malignancy.

We quantitatively evaluate a novel neural network pattern recognition approach for characterization of diagnostically challenging breast lesions in contrast-enhanced dynamic breast MRI. Eighty-two women with 84 indeterminate mammographic lesions (BIRADS III-IV, 38/46 benign/malignant lesions confirmed by histopathology and follow-up, median lesion diameter 12mm) were examined by dynamic contrast-enhanced breast MRI. The temporal signal dynamics results in an intensity time-series for each voxel represented by a 6-dimensional feature vector. These vectors were clustered by minimal-free-energy Vector Quantization (VQ), which identifies groups of pixels with similar enhancement kinetics as prototypical time-series, so-called codebook vectors. For comparison, conventional analysis based on lesion-specific averaged signal-intensity time-courses was performed according to a standardized semi-quantitative evaluation score. For quantitative assessment of diagnostic accuracy, areas under ROC curves (AUC) were computed for both VQ and standard classification methods. VQ increased the diagnostic accuracy for classification between benign and malignant lesions, as confirmed by quantitative ROC analysis: VQ results (AUC=0.760) clearly outperformed the conventional evaluation of lesion-specific averaged time-series (AUC=0.693). Thus, the diagnostic benefit of neural network VQ for MR mammography analysis is quantitatively documented by ROC evaluation in a large data base of diagnostically challenging small focal breast lesions. VQ outperforms the conventional method w.r.t. diagnostic accuracy.

Bayesian neural network (BNN) with automatic relevance determination (ARD) priors has the ability to assess the relevance of each input feature during network training. Our purpose is to investigate the potential use of BNN-with-ARD-priors for joint feature selection and classification in computer-aided diagnosis (CAD) of medical imaging. With ARD priors, each group of weights that connect an input feature to the hidden units is associated with a hyperparameter controlling the magnitudes of the weights. The hyperparameters and the weights are updated simultaneously during neural network training. A smaller hyperparameter will likely result in larger weight values and the corresponding feature will likely be more relevant to the output, and thus, to the classification task. For our study, a multivariate normal feature space is designed to include one feature with high classification performance in terms of both ideal observer and linear observer, two features with high ideal observer performance but low linear observer performance and 7 useless features. An exclusive-OR (XOR) feature space is designed to include 2 XOR features and 8 useless features. Our simulation results show that the ARD-BNN approach has the ability to select the optimal subset of features on the designed nonlinear feature spaces on which the linear approach fails. ARD-BNN has the ability to recognize features that have high ideal observer performance. Stepwise linear discriminant analysis (SWLDA) has the ability to select features that have high linear observer performance but fails to select features that have high ideal observer performance and low linear observer performance. The cross-validation results on clinical breast MRI data show that ARD-BNN yields statistically significant better performance than does the SWLDA-LDA approach. We believe that ARD-BNN is a promising method for pattern recognition in computer-aided diagnosis of medical imaging.

The goal of interactive search-assisted diagnosis (ISAD) is to enable doctors to make more informed decisions about a given case by providing a selection of similar annotated cases. For instance, a radiologist examining a suspicious mass could study labeled mammograms with similar conditions and weigh the outcome of their biopsy results before determining whether to recommend a biopsy. The fundamental challenge in developing ISAD systems is the identification of similar cases, not simply in terms of superficial image characteristics, but in a medically-relevant sense. This task involves three aspects: extraction of a representative set of features, identifying an appropriate measure of similarity in the high-dimensional feature space, and return the most similar matches at interactive speed. The first has been an active research area for several decades. The second has largely been ignored by the medical imaging community. The third can be achieved using the Diamond framework, an open-source platform that enables efficient exploration of large distributed complex data repositories. This paper focuses on the second aspect. We show that the choice of distance metric affects the accuracy of an ISAD system and that machine learning enables the construction of effective domain-specific distance metrics. In the learned distance, data points with the same labels (e.g., malignant masses) are closer than data points with different labels (e.g., malignant vs. benign). Thus, the labels of the near neighbors of a new case are likely to be informative. We present and evaluate several novel methods for distance metric learning and evaluate them on a database involving 2522 mass regions of interest (ROI) extracted from digital mammograms, with ground truth defined by biopsy results (1800 malignant, 722 benign). Our results show that learned distance metrics improve both classification (ROC curve) and retrieval performance.

Presentation of images with known pathology similar to that of a new unknown lesion would be helpful for radiologists in their diagnosis of breast cancer. In order to find images that are really similar and useful to radiologists, we determined the radiologists' subjective similarity ratings for pairs of masses, and investigated objective similarity measures that would agree well with the subjective ratings. Fifty sets of images, each of which included one image in the center and six other images to be compared with the center image, were selected; thus, 300 pairs of images were prepared. Ten breast radiologists provided the subjective similarity ratings for each image pair in terms of the overall impression for diagnosis. The objective similarity measures based on cross-correlation of the images, differences in feature values, and psychophysical measures by use of an artificial neural network were determined. The objective measures based on the cross-correlation were found to be not correlated with the subjective similarity ratings (r < 0.1). The differences in the features characterizing the margin were relatively strong indicators of the similarity (r > 0.40). When several image features were used, the differences-based objective measure was moderately correlated (r = 0.59) with the subjective ratings. The relatively high correlation coefficient (r = 0.74) was obtained for the psychophysical similarity measure. The similar images selected by use of the psychophysical measure can be useful to radiologists in the diagnosis of breast cancer.

In this paper, we compare two state-of-the-art classification techniques characterizing masses as either benign or malignant, using a dataset consisting of 271 cases (131 benign and 140 malignant), containing both a MLO and CC view. For suspect regions in a digitized mammogram, 12 out of 81 calculated image features have been selected for investigating the classification accuracy of support vector machines (SVMs) and Bayesian networks (BNs). Additional techniques for improving their performance were included in their comparison: the Manly transformation for achieving a normal distribution of image features and principal component analysis (PCA) for reducing our high-dimensional data. The performance of the classifiers were evaluated with Receiver Operating Characteristics (ROC) analysis. The classifiers were trained and tested using a k-fold cross-validation test method (k=10). It was found that the area under the ROC curve (A_z) of the BN increased significantly (p=0.0002) using the Manly transformation, from A_z = 0.767 to A_z = 0.795. The Manly transformation did not result in a significant change for SVMs. Also the difference between SVMs and BNs using the transformed dataset was not statistically significant (p=0.78). Applying PCA resulted in an improvement in classification accuracy of the naive Bayesian classifier, from A_z = 0.767 to A_z = 0.786. The difference in classification performance between BNs and SVMs after applying PCA was small and not statistically significant (p=0.11).

The surgical treatment of femur fractures, which often result from osteoporosis, is highly dependent on the quality of the femoral bone. Unsatisfying results of surgical interventions like early loosening of implants may be one result of altered bone quality. However, clinical diagnostic techniques to quantify local bone quality are limited and often highly observer dependent. Therefore, the development of tools, which automatically and reproducibly place regions of interest (ROI) and asses the local quality of the femoral bone in these ROIs would be of great help for clinicians. For this purpose, a method to position and deform ROIs automatically and reproducibly depending on the size and shape of the femur will be presented. Moreover, an approach to asses the femur quality, which is based on calculating texture features using co-occurrence matrices and these adaptive regions, will be proposed. For testing purposes, 15 CT-datasets of anatomical specimen of human femora are used. The correlation between the texture features and biomechanical properties of the proximal femoral bone is calculated. First results are very promising and show high correlation between the calculated features and biomechanical properties. Testing the method on a larger data pool and refining the algorithms to further increase its sensitivity for altered bone quality will be the next steps in this project.

Periprosthetic osteolysis is a disease triggered by the body's response to tiny wear fragments from total hip replacements (THR), which leads to localized bone loss and disappearance of the trabecular bone texture. We have been investigating methods of temporal radiographic texture analysis (tRTA) to help detect periprosthetic osteolysis. One method involves merging feature measurements at multiple time points using an LDA or BANN. The major drawback of this method is that several cases do not meet the inclusion criteria because of missing data, i.e., missing image data at the necessary time intervals. In this research, we investigated imputation methods to fill in missing data points using feature averaging, linear interpolation, and first and second order polynomial fitting. The database consisted of 101 THR cases with full data available from four follow-up intervals. For 200 iterations, missing data were randomly created to simulate a typical THR database, and the missing points were then filled in using the imputation methods. ROC analysis was used to assess the performance of tRTA in distinguishing between osteolysis and normal cases for the full database and each simulated database. The calculated values from the 200 iterations showed that the imputation methods produced negligible bias, and substantially decreased the variance of the AUC estimator, relative to excluding incomplete cases. The best performing imputation methods were those that heavily weighted the data points closest to the missing data. The results suggest that these imputation methods appear to be acceptable means to include cases with missing data for tRTA.

A computer aided root lesion detection method for digital dental X-rays is proposed using level set and complex wavelets. The detection method consists of two stages: preprocessing and root lesion detection. During preprocessing, a level set segmentation is applied to separate the teeth from the background. Tailored for the dental clinical environment, a segmentation clinical acceleration scheme is applied by using a support vector machine (SVM) classifier and individual principal component analysis (PCA) to provide an initial contour. Then, based on the segmentation result, root lesion detection is performed. Firstly, the teeth are isolated by the average intensity profile. Secondly, a center-line zero crossing based candidate generation is applied to generate the possible root lesion areas. Thirdly, the Dual-Tree Complex Wavelets Transform (DT-CWT) is used to further remove false positives. Lastly when the root lesion is detected, the area of root lesion is automatically marked with color indication representing different levels of seriousness. 150 real dental X-rays with various degrees of root lesions are used to test the proposed method. The results were validated by the dentist. Experimental results show that the proposed method is able to successfully detect the root lesion and provide visual assistance to the dentist.

Vertebral fractures are the most common osteoporosis-related fractures. It is important to detect vertebral fractures, because they are associated with increased risk of subsequent fractures, and because pharmacologic therapy can reduce the risk of subsequent fractures. Although vertebral fractures are often not clinically recognized, they can be visualized on lateral chest radiographs taken for other purposes. However, only 15-60% of vertebral fractures found on lateral chest radiographs are mentioned in radiology reports. The purpose of this study was to develop a computerized method for detection of vertebral fractures on lateral chest radiographs in order to assist radiologists' image interpretation. Our computerized method is based on the automated identification of upper and lower vertebral edges. In order to develop the scheme, radiologists provided morphometric data for each identifiable vertebra, which consisted of six points for each vertebra, for 25 normals and 20 cases with severe fractures. Anatomical information was obtained from morphometric data of normal cases in terms of vertebral heights, heights of vertebral disk spaces, and vertebral centerline. Computerized detection of vertebral fractures was based on the reduction in the heights of fractured vertebrae compared to adjacent vertebrae and normal reference data. Vertebral heights from morphometric data on normal cases were used as reference. On 138 chest radiographs (20 with fractures) the sensitivity of our method for detection of fracture cases was 95% (19/20) with 0.93 (110/118) false-positives per image. In conclusion, the computerized method would be useful for detection of potentially overlooked vertebral fractures on lateral chest radiographs.

Localization of catheter tips is the most common task in intensive care unit imaging. In this work, catheters appearing in digital chest radiographs acquired by portable chest x-rays were tracked using a semi-automatic method. Due to the fact that catheters are synthetic objects, its profile does not vary drastically over its length. Therefore, we use forward looking registration with normalized cross-correlation in order to take advantage of a priori information of the catheter profile. The registration is accomplished with a two-dimensional template representative of the catheter to be tracked generated using two seed points given by the user. To validate catheter tracking with this method, we look at two metrics: accuracy and precision. The algorithms results are compared to a ground truth established by catheter midlines marked by expert radiologists. Using 12 objects of interest comprised of naso-gastric, endo-tracheal tubes, and chest tubes, and PICC and central venous catheters, we find that our algorithm can fully track 75% of the objects of interest, with a average tracking accuracy and precision of 85.0%, 93.6% respectively using the above metrics. Such a technique would be useful for physicians wishing to verify the positioning of catheter tips using chest radiographs.

This study investigated the agreement between breast mass spiculation levels as rated subjectively by observers and a computer scheme. An image dataset with 1,263 mass regions was selected. First, three experienced observers independently rated subjectively the visualized spiculation levels of these mass regions and classified them into three categories (none/minimal, moderate, and severe/significant). We then developed a computerized scheme to detect mass margins and classify the spiculation levels of the suspected mass regions. The scheme applied a hybrid region growth algorithm to segment the mass regions. An edge map was computed inside a 30-pixel-wide band surrounding the mass boundary contour. The scheme then applied a threshold to convert the edge map into a binary image followed by labeling and detecting line orientation. In the original edge map the scheme computed the average local pixel value fluctuation. In the binary edge map, the scheme computed the ratio between the number of "spiculated" pixels and the number of total pixels inside the band. The scheme also computed mass region conspicuity using the original image. Using these three features, a Bayesian Belief Network (BBN) was built to classify mass regions into one of the three spiculation categories. We compared the inter-observer variation as well as agreement levels between the subjective and computerized ratings. Agreement rates between paired observers ranged from 41.3% to 58.8% (Kappa = 0.136 to 0.309). The agreement between the computer scheme and observers' average rating was 49.2% (Kappa = 0.218). This study demonstrated a large inter-observer variability in subjective rating of mass speculation levels as well as a large difference between the rating results of a computerized scheme and observers. As a result, in an Interactive Computer-Aided Diagnosis (ICAD) environment, CAD-selected reference regions may be considered "very similar" by some observers and "not similar" by others. Hence, improving the selection of actually visually similar reference regions by a computerized scheme remains an important yet unsolved task for ICAD development.

Radiologists can use the differences between the left and right breasts, or asymmetry, in mammograms to help detect certain malignant breast cancers. An image similarity method has been improved to make use of this knowledge base to recognize breast cancer. Image similarity is determined using computer-aided detection (CAD) prompts as the features, and then a cluster comparison is done to determine whether there is asymmetry. We develop the analysis through a combination of clustering and supervised learning of model parameters. This process correctly classifies cancerous mammograms 95% of the time, and all mammograms 84% of the time, and thus asymmetry is a measure that can play an important role in significantly improving computer-aided breast cancer detection systems. This technique represents an improvement in accuracy of 121% over commercial techniques on non-cancerous cases. Most computer-aided detection (CAD) systems are tested on images which contain cancer on the assumption that images without cancer would produce the same number of false positives. However, a pre-screening system is designed to remove the normal cases from consideration, and so the inclusion of a pre-screening system into CAD dramatically reduces the number of false positives reported by the CAD system. We define three methods for the inclusion of pre-screening into CAD, and improve the performance of the CAD system by over 70% at low levels of false positives.

Correction of patient motion is a fundamental preprocessing step for dynamic contrast-enhanced (DCE) breast MRI, removing artifacts induced by involuntary movement and facilitating quantitative analysis of contrast agent kinetics. Image registration algorithms commonly employed for this task align subsequent temporal images of the dynamic MRI by maximizing intensity-, correlation- or entropy-based similarity measures between image pairs. To compensate for global patient motion, frequently an initial affine linear or rigid transformation is estimated. Subsequently, local image variablity is reduced by maximizing local similarity measures and using viscous fluid or elastic regularization terms. We present a novel iterative scheme combining local and global registration into one single algorithm, limiting computational overhead, reducing interpolation artifacts and generally improving the quality of registration results. The relation between local and global motion is adjusted by the introduction of corresponding flexible weighting functions, allowing for a sound combination of both registration types and a potentially wider range of computable transformations. The proposed method is evaluated on both synthetic images and clinical breast MRI data. The results demonstrate that our method works stable and reliably compensates for common motion artifacts typical to DCE MR mammography.

The automatic segmentation of relevant structures such as skin edge, chest wall, or nipple in dynamic contrast enhanced MR imaging (DCE MRI) of the breast provides additional information for computer aided diagnosis (CAD) systems. Automatic reporting using BI-RADS criteria benefits of information about location of those structures. Lesion positions can be automatically described relatively to such reference structures for reporting purposes. Furthermore, this information can assist data reduction for computation expensive preprocessing such as registration, or for visualization of only the segments of current interest. In this paper, a novel automatic method for determining the air-breast boundary resp. skin edge, for approximation of the chest wall, and locating of the nipples is presented. The method consists of several steps which are built on top of each other. Automatic threshold computation leads to the air-breast boundary which is then analyzed to determine the location of the nipple. Finally, results of both steps are starting point for approximation of the chest wall. The proposed process was evaluated on a large data set of DCE MRI recorded by T1 sequences and yielded reasonable results in all cases.

The comparison of left and right mammograms is a common technique used by radiologists for the detection and diagnosis of masses. In mammography, computer-aided detection (CAD) schemes using bilateral subtraction technique have been reported. However, in breast ultrasonography, there are no reports on CAD schemes using comparison of left and right breasts. In this study, we propose a scheme of false positive reduction based on bilateral subtraction technique in whole breast ultrasound images. Mass candidate regions are detected by using the information of edge directions. Bilateral breast images are registered with reference to the nipple positions and skin lines. A false positive region is detected based on a comparison of the average gray values of a mass candidate region and a region with the same position and same size as the candidate region in the contralateral breast. In evaluating the effectiveness of the false positive reduction method, three normal and three abnormal bilateral pairs of whole breast images were employed. These abnormal breasts included six masses larger than 5 mm in diameter. The sensitivity was 83% (5/6) with 13.8 (165/12) false positives per breast before applying the proposed reduction method. By applying the method, false positives were reduced to 4.5 (54/12) per breast without removing a true positive region. This preliminary study indicates that the bilateral subtraction technique is effective for improving the performance of a CAD scheme in whole breast ultrasound images.

Medical information is evolving towards more complex multimedia data representation, as new imaging modalities are made available by sophisticated devices. Features such as segmented lesions can now be extracted through analysis techniques and need to be integrated into clinical patient data. The management of structured information extracted from multimedia has been addressed in knowledge based annotation systems providing methods to attach interpretative semantics to multimedia content. Building on these methods, we develop a new clinical imaging annotation system for computer aided breast cancer screening. The proposed system aims at more consistent, efficient and standardised data mark-up of digital and digitalised radiology images. The objective is to provide detailed characterisation of abnormalities as an aid in the diagnostic task through integrated annotation management. The system combines imaging analysis results and radiologist diagnostic information about suspicious findings by mapping well-established visual and low-level descriptors into pathology specific profiles. The versatile characterisation allows differentiating annotation descriptors for different types of findings. Our approach of semi-automatic integrated annotations supports increased quality assurance in screening practice. This is achieved through detailed and objective patient imaging information while providing user-friendly means for their manipulation that is oriented to relieving the radiologist's workload.

In this paper we present a novel approach for mass contour detection for 3D computer-aided detection (CAD) in digital breast tomosynthesis (DBT) data-sets. A hybrid active contour model, working directly on the projected views, is proposed. The responses of a wavelet filter applied on the projections are thresholded and combined to obtain markers for mass candidates. The contours of markers are extracted and serve as initialization for the active contour model, which is then used to extract mass contours in DBT projection images. A hybrid model is presented, taking into account several image-based external forces and implemented using a level-set formulation. A feature vector is computed from the detected contour, which may serve as input to a dedicated classifier. The segmentation method is applied to simulated images and to clinical cases. Image segmentation results are presented and compared to two standard active contour models. Evaluation of the performance on clinical data is obtained by comparison to manual segmentation by an expert. Performance on simulated images and visual performance assessment provide further illustration of the performance of the presented approach.

A method for computer-aided detection (CAD) of mammographic masses is proposed and a prototype CAD system is presented. The method is based on content-based image retrieval (CBIR). A mammogram database containing 2000 mammographic regions is built in our prototype CBIR-CAD system. Every region of interested (ROI) in the database has known pathology. Specifically, there are 583 ROIs depicting biopsy-proven masses, and the rest 1417 ROIs are normal. Whenever a suspicious ROI is detected in a mammogram by a radiologist, it can be submitted as a query to this CBIRCAD system. As the query results, a series of similar ROI images together with their known pathology knowledge will be retrieved from the database and displayed in the screen in descending order of their similarities to the query ROI to help the radiologist to make the diagnosis decision. Furthermore, our CBIR-CAD system will output a decision index (DI) to quantitatively indicate the probability that the query ROI contains a mass. The DI is calculated by the query matches. In the querying process, 24 features are extracted from each ROI to form a 24-dimensional vector. Euclidean distance in the 24-dimensional feature vector space is applied to measure the similarities between ROIs. The prototype CBIR-CAD system is evaluated based on the leave-one-out sampling scheme. The experiment results showed that the system can achieve a receiver operating characteristic (ROC) area index A_Z =0.84 for detection of mammographic masses, which is better than the best results achieved by the other known mass CAD systems.

In this paper, we present a fast method for microcalcification detection in Digital Breast Tomosynthesis. Instead of applying the straight-forward reconstruction/filtering/thresholding approach, the filtering is performed on projections before simple back-projection reconstruction. This leads to a reduced computation time since the number of projections is generally much smaller than the number of slices. For an average breast thickness and a typical number of projections, the number of operations is reduced by a factor in the range of 2 to 4. At the same time, the approach yields a negligible decrease of the contrast to noise ratio in the reconstructed slices. Image segmentation results are presented and compared to the previous method as visual performance assessment.

Multiplatform application of CAD systems in mammography is often limited due to image preprocessing steps that are tailored to the acquisition protocol such as the digitizer. The purpose of this study was to validate our knowledge-based CAD system across two different digitizers. Our system relies on the similarity of a query image with known cases stored in a knowledge database. Image similarity is assessed using information theory, without any image preprocessing. Therefore, we hypothesize that our CAD system can operate robustly across digitizers. We tested the hypothesis using two different datasets of mammographic regions of interest (ROIs) for mass detection. The two databases consisted of 1,820 and 1,809 ROIs extracted from DDSM mammograms digitized using a Lumisys and a Howtek scanner respectively. Three experiments were performed. First, we evaluated the CAD system on each dataset independently. Then, we evaluated the system on each dataset when the other dataset was used as the knowledge database. Finally, we assessed the CAD detection performance when the knowledge database contained mixed cases. Our CAD system had similar performance across digitizers (A_z=0.87±0.01 for Lumisys vs. A_z=0.8±0.01 for Howtek) when assessed independently. When the system was tested on one dataset while the other was used as the knowledge database, ROC performance declined marginally, mainly based on the partial ROC area index. This result suggests that blind translation of the system without some experience with cases digitized with the same digitizer is not recommended when the system is expected to operate at high sensitivity decision thresholds. When the system operated with a knowledge database of mixed cases, its performance across digitizers was robust yet slightly inferior to what observed independently.

The purpose of this study is to develop a Content-Based Image Retrieval (CBIR) system for mammographic computer-aided diagnosis. We have investigated the potential of using shape, texture, and intensity features to categorize masses that may lead to sorting similar image patterns in order to facilitate clinical viewing of mammographic masses. Experiments were conducted within a database that contains 243 masses (122 benign and 121 malignant). The retrieval performances using the individual feature was evaluated, and the best precision was determined to be 79.9% when using the curvature scale space descriptor (CSSD). By combining several selected shape features for retrieval, the precision was found to improve to 81.4%. By combining the shape, texture, and intensity features together, the precision was found to improve to 82.3%.

The contemporary goals of breast cancer treatment are not limited to cure but include maximizing quality of life. All breast cancer treatment can adversely affect breast appearance. Developing objective, quantifiable methods to assess breast appearance is important to understand the impact of deformity on patient quality of life, guide selection of current treatments, and make rational treatment advances. A few measures of aesthetic properties such as symmetry have been developed. They are computed from the distances between manually identified fiducial points on digital photographs. However, this is time-consuming and subject to intra- and inter-observer variability. The purpose of this study is to investigate methods for automatic localization of fiducial points on anterior-posterior digital photographs taken to document the outcomes of breast reconstruction. Particular emphasis is placed on automatic localization of the nipple complex since the most widely used aesthetic measure, the Breast Retraction Assessment, quantifies the symmetry of nipple locations. The nipple complexes are automatically localized using normalized cross-correlation with a template bank of variants of Gaussian and Laplacian of Gaussian filters. A probability map of likely nipple locations determined from the image database is used to reduce the number of false positive detections from the matched filter operation. The accuracy of the nipple detection was evaluated relative to markings made by three human observers. The impact of using the fiducial point locations as identified by the automatic method, as opposed to the manual method, on the calculation of the Breast Retraction Assessment was also evaluated.

We propose a methodological framework for texture analysis in medical images that is based on Vector Quantization (VQ), a method traditionally used for image compression. In this framework, the codeword usage histogram is used as a texture descriptor of the image. This descriptor can be used effectively for similarity searches, clustering, classification and other retrieval operations. We present an application of this approach to the analysis of x-ray galactograms; we analyze the texture in retroareolar regions of interests (ROIs) in order to distinguish between patients with reported galactographic findings and normal subjects. We decompose these ROIs into equi-size blocks and use VQ to represent each block with the closest codeword from a codebook. Each image is represented as a vector of frequencies of codeword appearance. We perform k-nearest neighbor classification of the texture patterns employing the histogram model as a similarity measure. The classification accuracy reached up to 96% for certain experimental settings; these results demonstrate that the proposed approach can be effective in performing similarity analysis of texture patterns in breast imaging. The proposed texture analysis framework has a potential to assist the interpretation of clinical images in general and facilitate the investigation of relationships among structure, texture and function or pathology.

In 2006, breast cancer is expected to continue as the leading form of cancer diagnosed in women, and the second leading cause of cancer mortality in this group. A method that has proven useful for guiding the choice of treatment strategy is the assessment of histological tumor grade. The grading is based upon the mitosis count, nuclear pleomorphism, and tubular formation, and is known to be subject to inter-observer variability. Since cancer grade is one of the most significant predictors of prognosis, errors in grading can affect patient management and outcome. Hence, there is a need to develop a breast cancer-grading tool that is minimally operator dependent to reduce variability associated with the current grading system, and thereby reduce uncertainty that may impact patient outcome. In this work, we explored the potential of a computer-based approach using fractal analysis as a quantitative measure of cancer grade for breast specimens. More specifically, we developed and optimized computational tools to compute the fractal dimension of low- versus high-grade breast sections and found them to be significantly different, 1.3±0.10 versus 1.49±0.10, respectively (Kolmogorov-Smirnov test, p<0.001). These results indicate that fractal dimension (a measure of morphologic complexity) may be a useful tool for demarcating low- versus high-grade cancer specimens, and has potential as an objective measure of breast cancer grade. Such prognostic value could provide more sensitive and specific information that would reduce inter-observer variability by aiding the pathologist in grading cancers.

Computer aided detection (CADe) systems often present multiple false-positives per image in projection mammography due to overlapping anatomy. To reduce the number of such false-positives, we propose performing CADe on image pairs acquired using a bi-plane correlation imaging (BCI) technique. In this technique, images are acquired of each breast at two different projection angles. A traditional CADe algorithm operates on each image to identify suspected lesions. The suspicious areas from both projections are then geometrically correlated, eliminating any lesion that is not identified on both views. Proof of concept studies showed that that the BCI technique reduced the numbers of false-positives per case up to 70%.

Women with dense breasts have an increased risk of breast cancer. Breast density is typically measured as the percent density (PD), the percentage of non-fatty (i.e., dense) tissue in breast images. Mammographic PD estimates vary, in part, due to the projective nature of mammograms. Digital breast tomosynthesis (DBT) is a novel radiographic method in which 3D images of the breast are reconstructed from a small number of projection (source) images, acquired at different positions of the x-ray focus. DBT provides superior visualization of breast tissue and has improved sensitivity and specificity as compared to mammography. Our long-term goal is to test the hypothesis that PD obtained from DBT is superior in estimating cancer risk compared with other modalities. As a first step, we have analyzed the PD estimates from DBT source projections since the results would be independent of the reconstruction method. We estimated PD from MLO mammograms (PD_M) and from individual DBT projections (PD_T). We observed good agreement between PD_M and PD_T from the central projection images of 40 women. This suggests that variations in breast positioning, dose, and scatter between mammography and DBT do not negatively affect PD estimation. The PD_T estimated from individual DBT projections of nine women varied with the angle between the projections. This variation is caused by the 3D arrangement of the breast dense tissue and the acquisition geometry.

We are developing a computer based aid for automated analysis of digital mammograms and digital breast tomosynthesis (DBT) images. The ultimate goal is to establish correspondence between regions in images obtained using different modalities. This paper is focused on establishing point correspondences between mammograms and DBT projections. Correspondence has been established utilizing two similarity criteria, one based on topology of prominent elongated anatomical structures, and another based on texture near the potential correspondence points. We evaluated robustness of the described technique with respect to variations in x-ray tube angle. The evaluation included 72 image pairs: 9 DBT projections from both left and right breasts of 4 women and corresponding mammograms. The evaluation was performed by a consensus between two trained observers. Two images with highlighted pairs of automatically established point correspondences were presented. The observers were asked to manually identify the correct correspondences and measure the displacement error. Points for which correspondences could not be manually identified were excluded. The topology method automatically generated an average of 12.2 correspondences per image pair, for which the average measured displacement error was 1.33 mm (N = 10.5). The texture method generated 18.6 correspondences with an average measured error of 1.80 mm (N = 14.7). The algorithms were found to be robust; the number of correspondences and the average displacement did not significantly change with variations in tube angle.

Purpose: The efficiency of the detection of pulmonary nodules by a radiologist with the help of CAD is influenced by the user interface of the system. Marker with a visually dominant appearance may distract the radiologist from other parts of the screen. Purpose was to analyse the influence of different CAD markers on radiologist's performance. Materials and methods: 10 radiologists analysed 150 pictures of chest CT slices. Every picture contained a CAD marker; five different types of markers were used - each respectively on 30 pictures (1: thick walled square, 2: thin walled circle, 3: small arrow, 4: pixel sized point on nodule, 5: very subtle change of colour). One hundred images contained one nodule: CAD markers marked this finding in 50 cases; in 50 cases a false positive finding was marked instead. The remaining 50 images contained no nodule but a marker of a false positive finding. The radiologists had to decide for each image if there was a nodule visible and either click on the nodule or on a button "no finding". Sensitivity and specificity were calculated for each marker type. Results: Mean sensitivity was 59%, 62%, 64%, 65% and 64% for marker 1 to 5, respectively. Specificity was 50%, 51%, 64%, 45% and 67%. In the cases with false positive findings sensitivity for detection of the unmarked nodule was 41%, 58%, 59%, 49% and 54%. New work to be presented: The study shows that the marker type influences radiologist's sensitivity and distraction from other findings. Conclusion: Of the tested markers a small arrow was most efficient for the presentation of the results to the radiologist.

Computer-aided diagnostic (CAD) schemes have been developed for assisting radiologists in the detection of various lesions in medical images. The reliable evaluation of CAD schemes is an important task in the field of CAD research. In the past, many evaluation approaches, such as the resubstitution, leave-one-out, cross-valiation, and hold-out methods have been used for evaluating the performance of various CAD schemes. However, some important issues in the evaluation of CAD schemes have not been systematically analyzed, either theoretically or experimentally. The first important issue is the analysis and comparison of various evaluation methods in terms of some characteristics, in particular, the bias and the generalization performance of the trained CAD schemes. The second includes the analysis of pitfalls in the incorrect use of various evaluation methods and the effective approaches to the reduction of the bias and variance caused by these pitfalls. We attempt to address the above important issues in this article. We believe that this article would be useful to researchers in the field of CAD research for selecting appropriate evluation methods and for improving the reliability of the estimated performance of their CAD schemes.

The development of image databases for CAD research is not a trivial task. The collection and management of images and their related metadata from multiple sources is a time-consuming but necessary process. By standardizing and centralizing the methods in which these data are maintained, one can generate subsets of a larger database that match the specific criteria needed for a particular research project in a quick and efficient manner. A research-oriented management system of this type is highly desirable in a multi-modality CAD research environment. An online, webbased database system for the storage and management of research-specific medical image metadata was designed for use with four modalities of breast imaging: screen-film mammography, full-field digital mammography, breast ultrasound and breast MRI. The system was designed to consolidate data from multiple clinical sources and provide the user with the ability to anonymize the data. Input concerning the type of data to be stored as well as desired searchable parameters was solicited from researchers in each modality. The backbone of the database was created using MySQL. A robust and easy-to-use interface for entering, removing, modifying and searching information in the database was created using HTML and PHP. This standardized system can be accessed using any modern web-browsing software and is fundamental for our various research projects on computer-aided detection, diagnosis, cancer risk assessment, multimodality lesion assessment, and prognosis. Our CAD database system stores large amounts of research-related metadata and successfully generates subsets of cases that match the user's desired search criteria.

The development of computer-aided diagnostic (CAD) systems requires an initial establishment of "truth" by expert human observers. Potential inconsistencies in the "truth" data must be identified and corrected before investigators can rely on this data. We developed a quality assurance model to supplement the "truth" collection process for lung nodules on CT scans. A two-phase process was established for the interpretation of CT scans by four radiologists. During the initial "blinded read," radiologists independently assigned lesions they identified into one of three categories: "nodule ⩾ 3mm," "nodule < 3mm," or "non-nodule ⩾ 3mm." During the subsequent "unblinded read," the blinded read results of all radiologists were revealed. The radiologists then independently reviewed their marks along with their colleague's marks; a radiologist's own marks could be left unchanged, deleted, switched in terms of lesion category, or additional marks could be added. The final set of marks underwent quality assurance, which consisted of identification of potential errors that occurred during the reading process and error correction. All marks were visually grouped into discrete nodules. Six categories of potential error were defined, and any nodule with a mark that satisfied the criterion for one of these categories was referred to the radiologist who assigned the mark in question. The radiologist either corrected the mark or confirmed that the mark was intentional. A total of 829 nodules were identified by at least one radiologist in 100 CT scans through the two-phase process designed to capture "truth." The quality assurance process yielded 81 nodules with potential errors. The establishment of "truth" must incorporate a quality assurance model to guarantee the integrity of the "truth" that will provide the basis for the training and testing of CAD systems.

The Lung Image Database Consortium (LIDC) has provided a publicly available collection of CT images with nodule markings from four radiologists. The LIDC protocol does not require radiologists to reach a consensus during the reading process, and as a result, there are varying levels of reader agreement for each potential nodule with no explicit reference standard for nodules. The purpose of this work was to investigate the effects of the level of reader agreement on the development of a reference standard and the subsequent impact on CAD performance. Ninety series were downloaded from the LIDC database. Four different reference standards were created based on the markings of the LIDC radiologists, reflecting four different levels of reader agreement. All series were analyzed with a research CAD system and its performance was measured against each of the four standards. Between the standards with the lowest (any 1 of 4 readers) and highest (all 4 readers) required level of reader agreement, the number of nodules ⩾ 3 mm decreased 48% (from 174 to 90) and CAD sensitivity for nodules ⩾ 3 mm increased from 0.70 ± 0.34 to 0.79 ± 0.35. Between the same reference standards, the number of nodules < 3 mm decreased 84% (from 483 to 75) and CAD sensitivity for nodules < 3 mm increased from 0.30 ± 0.29 to 0.51 ± 0.45. This research illustrates the importance of indicating the method used to form the reference standard, since the method influences both the number of nodules and reported CAD performance.

Magnetic resonance image (MRI) has provided an imageological support into the clinical diagnosis and prediction of Alzheimer disease (AD) progress. Currently, the clinical use of MRI data on AD diagnosis is qualitative via visual inspection and less accurate. To provide assistance to physicians in improving the accuracy and sensitivity of the AD diagnose and the clinical outcome of the disease, we developed a computer-assisted analysis package that analyzed the MRI data of an individual patient in comparison with a group of normal controls. The package is based on the principle of the well established and widely used voxel-based morphometry (VBM) and SPM software. All analysis procedure is automated and streamlined. With only one mouse-click, the whole procedure was finished within 15 minutes. With the interactive display and anatomical automatic labeling toolbox, the final result and report supply the brain regional structure difference, the quantitative assessment and visual inspections by physicians and scientific researcher. The brain regions which affected by AD are consonant in the main with the clinical diagnosis, which are reviewed by physicians. In result, the computer package provides physician with an automatic and assistant tool for prediction using MRI. This package could be valuable tool assisting physicians in making their clinical diagnosis decisions.

Magnetic resonance angiography (MRA) is routinely employed in the diagnosis of cerebrovascular disease. Unruptured aneurysms and arterial occlusions can be detected in examinations using MRA. This paper describes a computerized detection method of arterial occlusion in MRA studies. Our database consists of 100 MRA studies, including 85 normal cases and 15 abnormal cases with arterial occlusion. Detection of abnormality is based on comparison with a reference (normal) MRA study with all the vessel known. Vessel regions in a 3D target MRA study is first segmented by using thresholding and region growing techniques. Image registration is then performed so as to maximize the overlapping of the vessel regions in the target image and the reference image. The segmented vessel regions are then classified into eight arteries based on comparison of the target image and the reference image. Relative lengths of the eight arteries are used as eight features in classifying the normal and arterial occlusion cases. Classifier based on the distance of a case from the center of distribution of normal cases is employed for distinguishing between normal cases and abnormal cases. The sensitivity and specificity for the detection of abnormal cases with arterial occlusion is 80.0% (12/15) and 95.3% (81/85), respectively. The potential of our proposed method in detecting arterial occlusion is demonstrated.

Our purpose is in the automated evaluation of the physiological relevance of lesions in coronary angiograms. We aim to extract as much as possible quantitative information about the physiological condition of the heart from standard angiographic image sequences. Coronary angiography is still the gold standard for evaluating and diagnosing coronary abnormalities as it is able to locate precisely the coronary artery lesions. The dimensions of the stenosis can be assessed nowadays successfully with image processing based Quantitative Coronary Angiography (QCA) techniques. Our purpose is to assess the clinical relevance of the pertinent stenosis. We therefore analyze the myocardial perfusion as revealed in standard angiographic image sequences. In a Region-of-Interest (ROI) on the angiogram (without an overlaying major blood vessel) the contrast is measured as a function of time (the so-called time-density curve). The required hyperemic state of exercise is induced artificially by the injection of a vasodilator drug e.g. papaverine. In order to minimize motion artifacts we select based on the recorded ECG signal end-diastolic images in both a basal and a hyperemic run in the same projection to position the ROI. We present the development of the algorithms together with results of a small study of 20 patients which have been catheterized following the standard protocol.

Recently, multi-slice helical CT technology was developed. Unlike the conventional helical CT, we can obtain CT images of two or more slices with 1 time scan. Therefore, we can get many pictures with a clear contrast images and thin slice images in one time of scanning. The purpose of this presentation is to evaluate the proposed automatic extraction bronchus and pulmonary vein and artery on multi-slice CT images. The bronchus is extracted by application with region growing technique and the morphological filters, 3D distance transformation. These results indicate that the proposed algorithm provides the ability to accurately develop an automatic extraction algorithm of the bronchus on multi-slice CT images. In this report, we used pulmonary vein and artery marked by the doctor, It aims to discover an amount of the feature necessary for classifying the pulmonary vein and artery by using the anatomical feature. The classification of the pulmonary vein and artery is thought necessary information that it is state of tuber benign or malignity judgment. It is very important to separate the contact part of the lung blood vessel in classifying pulmonary vein and artery. Then, it aims to discover the feature of the contact part of the lung blood vessel in this report.

CT pulmonary angiography (CTPA) has been reported to be an effective means for clinical diagnosis of pulmonary embolism (PE). We are developing a computer-aided detection (CAD) system to assist radiologist in PE detection in CTPA images. 3D multiscale filters in combination with a newly designed response function derived from the eigenvalues of Hessian matrices is used to enhance vascular structures including the vessel bifurcations and suppress non-vessel structures such as the lymphoid tissues surrounding the vessels. A hierarchical EM estimation is then used to segment the vessels by extracting the high response voxels at each scale. The segmented vessels are pre-screened for suspicious PE areas using a second adaptive multiscale EM estimation. A rule-based false positive (FP) reduction method was designed to identify the true PEs based on the features of PE and vessels. 43 CTPA scans were used as an independent test set to evaluate the performance of PE detection. Experienced chest radiologists identified the PE locations which were used as "gold standard". 435 PEs were identified in the artery branches, of which 172 and 263 were subsegmental and proximal to the subsegmental, respectively. The computer-detected volume was considered true positive (TP) when it overlapped with 10% or more of the gold standard PE volume. Our preliminary test results show that, at an average of 33 and 24 FPs/case, the sensitivities of our PE detection method were 81% and 78%, respectively, for proximal PEs, and 79% and 73%, respectively, for subsegmental PEs. The study demonstrates the feasibility that the automated method can identify PE accurately on CTPA images. Further study is underway to improve the sensitivity and reduce the FPs.

We evaluated a Pareto front-based multi-objective evolutionary algorithm for optimizing our CT colonography (CTC) computer-aided detection (CAD) system. The system identifies colonic polyps based on curvature and volumetric based features, where a set of thresholds for these features was optimized by the evolutionary algorithm. We utilized a two-fold cross-validation (CV) method to test if the optimized thresholds can be generalized to new data sets. We performed the CV method on 133 patients; each patient had a prone and a supine scan. There were 103 colonoscopically confirmed polyps resulting in 188 positive detections in CTC reading from either the prone or the supine scan or both. In the two-fold CV, we randomly divided the 133 patients into two cohorts. Each cohort was used to obtain the Pareto front by a multi-objective genetic algorithm, where a set of optimized thresholds was applied on the test cohort to get test results. This process was repeated twice so that each cohort was used in the training and testing process once. We averaged the two training Pareto fronts as our final training Pareto front and averaged the test results from the two runs in the CV as our final test results. Our experiments demonstrated that the averaged testing results were close to the mean Pareto front determined from the training process. We conclude that the Pareto front-based algorithm appears to be generalizable to new test data.

Multiple classifiers working collaboratively can usually achieve better performance than any single classifier working independently. Our CT colonography computer-aided detection (CAD) system uses support vector machines (SVM) as the classifier. In this paper, we developed and evaluated two schemes to collaboratively apply multiple SVMs in the same CAD system. One is to put the classifiers in a sequence (SVM sequence) and apply them one after another; the other is to put the classifiers in a committee (SVM committee) and use the committee decision for the classification. We compared the sequence order (best-first, worst-first and random) in the SVM sequence and two decision functions in the SVM committee (majority vote and sum probability). The experiments were conducted on 786 CTC datasets, with 63 polyp detections. We used 10-fold cross validation to generate the FROC curves, and conducted 100 bootstraps to evaluate the performance variation. The result showed that collaborative classifiers performed much better than individual classifiers. The SVM sequence had slightly better accuracy than the SVM committee but also had bigger performance variation.

In this paper, we first introduce three different geometric features including shape index, curvedness and sphericity ratio, for colonic polyp detection. A new color coding scheme is designed to highlight the detected polyps, and help radiologists to distinguish them from other tissues more easily. The key idea is to place the detected polyp candidates at the same locations in a newly created polygonal dataset with exactly the same topological and geometrical properties as the triangulated mesh surface of real colon dataset, and assign different colors to the two separated datasets to highlight the polyps. Finally, we validate the proposed polyp detection framework and color coding scheme by computer simulated and real colon datasets. For sixteen synthetic polyps with different shapes and different sizes, the sensitivity is 100%, and false positive is 0.

Colorectal cancer is the second leading cause of cancer deaths, and ranks third for new cancer cases and cancer mortality for both men and women. However, its death rate can be dramatically reduced by appropriate treatment when early detection is available. The purpose of colonoscopy is to identify and assess the severity of lesions, which may be flat or protruding. Due to the subjective nature of the examination, colonoscopic proficiency is highly variable and dependent upon the colonoscopist's knowledge and experience. An automated image processing system providing an objective, rapid, and inexpensive analysis of video from a standard colonoscope could provide a valuable tool for screening and diagnosis. In this paper, we present the design, functionality and preliminary results of its Computer-Aided-Diagnosis (CAD) system for colonoscopy - ColonoCAD^TM. ColonoCAD is a complex multi-sensor, multi-data and multi-algorithm image processing system, incorporating data management and visualization, video quality assessment and enhancement, calibration, multiple view based reconstruction, feature extraction and classification. As this is a new field in medical image processing, our hope is that this paper will provide the framework to encourage and facilitate collaboration and discussion between industry, academia, and medical practitioners.

Hessian matrix is the square matrix of second partial derivatives of a scalar-valued function and is well known for object recognition in computer vision and medical shape analysis. Previous curvature based polyp detection algorithms generate myriad of false positives. Hessian-matrix based method, however, is more sensitive to local shape features, so easily reduce false positives. Calculation of Hessian matrix on 3D CT data and Eigen decomposition of the matrix gives three Eigen values and vectors at each voxel. Using these Eigen values, we can figure out which type of intensity structures (blob, line, and sheet-like) is on the given voxel. We focus on detecting blob-like object automatically. In the inner colonic wall structures, blob-like, line-like, and sheet-like objects represent polyps, folds and wall, respectively. In addition, to improve the performance of the algorithm, Gaussian blurring factor and shape threshold parameters are optimized. Before Hessian matrix calculation, smoothing the given region using Gaussian kernel with small deviation is necessary to enhance local intensity structures. To optimize the parameters and validate this method, we have produced anthropomorphic pig phantoms. Fourteen phantoms with 103 polyps (16 polyps <6mm, 87 >= 6mm) were used. CT scan was performed with 1mm slice thickness. Our detection algorithm found 84 polyps (81.6%) correctly. Average number of false positives is 7.9 at each CT scan. This results show that our algorithm is clinically applicable for polyp detection, because of high sensitivity and relatively low false positive detections.

This paper presents a new method for detecting colonic polyps from abdominal CT images based on Hessian matrix analysis. Recently, virtual colonoscopy (VC) has widely received attention as a new and less-invasive colon diagnostic method. A physician diagnoses the inside of the colon using a virtual colonoscopy system. However, since the colon has many haustra and its shape is long and convoluted, a physician has to change viewpoints and viewing directions of the virtual camera many times while diagnosing. Lesions behind haustra may be overlooked. Thus this paper proposes an automated colonic polyp detection method from 3D abdominal CT images. Colonic polyps are located on the colonic wall, and their CT values are higher than colonic lumen regions. In addition, CT values inside polyps tend to gradually increase from outward to inward (blob-like structure). We employ a blob structure enhancement filter based on the eigenvalues of a Hessian matrix to detect polyps with the above blob-shaped characteristics. For reducing FPs, we eliminate polyp candidate regions in which the maximum output value of the blob structure enhancement filter is smaller than given threshold values. Also small regions are removed from candidates. We applied the proposed method to 23 cases of abdominal CT images. Overall, 74.4% of the polyps were detected with 3.8 FPs per case.

This paper describes a method for detecting hemorrhages and exudates in ocular fundus images. The detection of hemorrhages and exudates is important in order to diagnose diabetic retinopathy. Diabetic retinopathy is one of the most significant factors contributing to blindness, and early detection and treatment are important. In this study, hemorrhages and exudates were automatically detected in fundus images without using fluorescein angiograms. Subsequently, the blood vessel regions incorrectly detected as hemorrhages were eliminated by first examining the structure of the blood vessels and then evaluating the length-to-width ratio. Finally, the false positives were eliminated by checking the following features extracted from candidate images: the number of pixels, contrast, 13 features calculated from the co-occurrence matrix, two features based on gray-level difference statistics, and two features calculated from the extrema method. The sensitivity of detecting hemorrhages in the fundus images was 85% and that of detecting exudates was 77%. Our fully automated scheme could accurately detect hemorrhages and exudates.

A central problem in the development of a mass-screening tool for atherosclerotic plaque is automatic calcification detection. The mass-screening aspect implies that the detection process should be fast and reliable. In this paper we present a first step in this direction by introducing a semi-automatic calcification classification tool based on non-linear stretching, an image enhancement method that focusses on local image statistics. The calcified areas are approximated by a coarse brush, which in our case is mimicked by taking the ground truth, provided by radiologists, and dilating it with circular structuring elements of varying sizes. Thresholds are then examined for the different structuring elements, that yield optimal results on the enhanced image. The results in this preliminary study which contains 19 images of varying calcification degree, fully annotated by medical experts, show a significant increase in accuracy when the methodology is validated on a region of interest containing the areas of a simulated coarse brush.

We developed a computer-aided diagnostic (CAD) scheme for assisting radiologists in the detection of asymmetric abnormalities on a single whole-body bone scintigram by applying a contralateral subtraction (CS) technique. Twenty whole-body bone scans including 107 abnormal lesions in anterior and/or posterior images (the number of lesions per case ranged from 1 to 16, mean 5.4) were used in this study. In our scheme, the original bone scan image was flipped horizontally to provide a mirror image. The mirror image was first rotated and shifted globally to match the original image approximately, and then was nonlinearly warped by use of an elastic matching technique in order to match the original image accurately. We applied a nonlinear lookup table to convert the difference in pixel values between the original and the warped images to new pixel values for a CS image, in order to enhance dark shadows at the locations of abnormal lesions where uptake of radioisotope was asymmetrically high, and to suppress light shadows of the lesions on the contralateral side. In addition, we applied a CAD scheme for the detection of asymmetric abnormalities by use of rule-based tests and sequential application of artificial neural networks with 25 image features extracted from the original and CS images. The performance of the CAD scheme, which was evaluated by a leave-one-case-out method, indicated an average sensitivity of 80.4 % with 3.8 false positives per case. This CAD scheme with the contralateral subtraction technique has the potential to improve radiologists' diagnostic accuracy and could be used for computerized identification of asymmetric abnormalities on whole-body bone scans.

Cervical Cancer is the second most common cancer among women worldwide and the leading cause of cancer mortality of women in developing countries. If detected early and treated adequately, cervical cancer can be virtually prevented. Cervical precursor lesions and invasive cancer exhibit certain morphologic features that can be identified during a visual inspection exam. Digital imaging technologies allow us to assist the physician with a Computer-Aided Diagnosis (CAD) system. In colposcopy, epithelium that turns white after application of acetic acid is called acetowhite epithelium. Acetowhite epithelium is one of the major diagnostic features observed in detecting cancer and pre-cancerous regions. Automatic extraction of acetowhite regions from cervical images has been a challenging task due to specular reflection, various illumination conditions, and most importantly, large intra-patient variation. This paper presents a multi-step acetowhite region detection system to analyze the acetowhite lesions in cervical images automatically. First, the system calibrates the color of the cervical images to be independent of screening devices. Second, the anatomy of the uterine cervix is analyzed in terms of cervix region, external os region, columnar region, and squamous region. Third, the squamous region is further analyzed and subregions based on three levels of acetowhite are identified. The extracted acetowhite regions are accompanied by color scores to indicate the different levels of acetowhite. The system has been evaluated by 40 human subjects' data and demonstrates high correlation with experts' annotations.

Primary malignant liver tumor, including hepatocellular carcinoma (HCC), caused 1.25 million deaths per year worldwide. Multiphase CT images offer clinicians important information about hepatic cancer. The presence of HCC is indicated by high-intensity regions in arterial phase images and low-intensity regions in equilibrium phase images following enhancement with contrast material. We propose an automatic method for detecting HCC based on edge detection and subtraction processing. Within a liver area segmented according to our scheme, black regions are selected by subtracting the equilibrium phase images to the corresponding registrated arterial phase images. From these black regions, the HCC candidates are extracted as the areas without edges by using Sobel and LoG edge detection filters. The false-positive (FP) candidates are eliminated by using six features extracted from the cancer and liver regions. Other FPs are further eliminated by opening processing. Finally, an expansion process is applied to acquire the 3D shape of the HCC. The cases used in this experiment were from the CT images of 44 patients, which included 44 HCCs. We extracted 97.7% (43/44) HCCs successfully by our proposed method, with an average number of 2.1 FPs per case. The result demonstrates that our edge-detection-based method is effective in locating the cancer region by using the information obtained from different phase images.

In Japan, stomach cancer is one of the three most common causes of death from cancer. Since periodic health checks of stomach X-rays have become more widely carried out, the physicians' burdens have been increasing in the mass screening to detect initial symptoms of a disease. For the purpose of automatic diagnosis, we try to develop a computer-assisted lesion detection system for stomach screening. The proposed system has two databases. One is the stomach shape database that consists of the computer graphics stomach 3D models based on biomechanics simulation and their projected 2D images. The other is the normal appearance database that is constructed by learning patterns in a normal patient training set. The stomach contour is extracted from an X-ray image including a barium filled region by the following steps. Firstly, the approximated stomach region is obtained by nonrigid registration based on mutual information. We define nonrigid transformation as one that includes translations, rotations, scaling, air-barium interface and weights of eigenvectors determined by principal components analysis in the stomach shape database. Secondly, the accurate stomach contour is extracted from the gradient of an image by using the Dynamic Programming. After then, stomach lesions are detected by inspecting whether the Mahalanobis distance from the mean in the normal appearance database is longer than a suitable value on the extracted stomach contour. We applied our system to 75 X-ray images of barium-filled stomach to show its validity.

Many cervical Computer-Aided Diagnosis (CAD) methods rely on measuring gradual appearance changes on the cervix after the application of a contrast agent. Image registration has been used to ensure pixel correspondence to the same tissue location throughout the whole temporal sequence but, to date, there is no reliable mean of testing its accuracy to compensate for patient and tissue movement. We present an independent system to use automatically extracted and matched features from a colposcopic image sequence in order to generate position landmarks. These landmarks may be used either to measure the accuracy of a registration method to align any pair of images from the colposcopic sequence or as a cue for registration. The algorithm selects sets of matched features that extend through the whole image sequence allowing to locate, in a reliable and unbiased way, a tissue point throughout the whole image sequence. Experiments on real colposcopy image sequences show that the approach is robust, reliable, and leads to geometrically coherent sets of landmarks that correspond to visually recognizable regions. We use the extracted landmarks to test the precision of some of the cervical registration algorithms previously presented in the literature.

The detection and aging of bruises is important within clinical and forensic environments. Traditionally, visual and photographic assessment of bruise color is used to determine age, but this substantially subjective technique has been shown to be inaccurate and unreliable. The purpose of this study was to develop a technique to spectrally-age bruises using a reflective multi-spectral imaging system that minimizes the filtering and hardware requirements while achieving acceptable accuracy. This approach will then be incorporated into a handheld, point-of-care technology that is clinically-viable and affordable. Sixteen bruises from elder residents of a long term care facility were imaged over time. A multi-spectral system collected images through eleven narrow band (~10 nm FWHM) filters having center wavelengths ranging between 370-970 nm corresponding to specific skin and blood chromophores. Normalized bruise reflectance (NBR)- defined as the ratio of optical reflectance coefficient of bruised skin over that of normal skin- was calculated for all bruises at all wavelengths. The smallest mean NBR, regardless of bruise age, was found at wavelength between 555 & 577nm suggesting that contrast in bruises are from the hemoglobin, and that they linger for a long duration. A contrast metric, based on the NBR at 460nm and 650nm, was found to be sensitive to age and requires further investigation. Overall, the study identified four key wavelengths that have promise to characterize bruise age. However, the high variability across the bruises imaged in this study complicates the development of a handheld detection system until additional data is available.

Multiple sclerosis (MS) is a progressive neurological disease affecting myelin pathways. MRI has become the medical imaging study of choice both for the diagnosis and for the follow-up and monitoring of multiple sclerosis. The progression of the disease is variable, and requires routine follow-up to document disease exacerbation, improvement, or stability of the characteristic MS lesions or plaques. The difficulties with using MRI as a monitoring tool are the significant quantities of time needed by the radiologist to actually measure the size of the lesions, and the poor reproducibility of these manual measurements. A CAD system for automatic image analysis improves clinical efficiency and standardizes the lesion measurements. Multiple sclerosis is a disease well suited for automated analysis. The segmentation algorithm devised classifies normal and abnormal brain structures and measures the volume of multiple sclerosis lesions using fuzzy c-means clustering with incorporated spatial (sFCM) information. First, an intracranial structures mask in T1 image data is localized and then superimposed in FLAIR image data. Next, MS lesions are identified by sFCM and quantified within a predefined volume. The initial validation process confirms a satisfactory comparison of automatic segmentation to manual outline by a neuroradiologist and the results will be presented.