This PDF file contains the front matter associated with SPIE Proceedings Volume 7624, including the Title Page, Copyright information, Table of Contents, and the Conference Committee listing.

Reviewing videos from medical procedures is a tedious work that requires concentration for extended hours and usually screens thousands of frames to find only a few positive cases that indicate probable presence of disease. Computational classification algorithms are sought to automate the reviewing process. The class imbalance problem becomes challenging when the learning process is driven by relative few minority class samples. The learning algorithms using imbalanced data sets generally result in large number of false negatives. In this article, we present an efficient rebalancing method for finding video frames that contain bleeding lesions. The majority class generally has clusters of data within them. Here we cluster the majority class and under-sample the each cluster based on its variance so that useful examples would not be lost during the under-sampling process. The balance of bleeding to non-bleeding frames is restored by the proposed cluster-based under-sampling and oversampling using Synthetic Minority Over-sampling Technique (SMOTE). Experiments were conducted using synthetic data and videos manually annotated by medical specialists for obscure bleeding detection. Our method achieved a high average sensitivity and specificity.

The evaluation of automated classifiers in computer-aided diagnosis of medical images often involves a training dataset for classifier design and a test dataset for performance estimation in terms of, e.g., area under the receiver operating characteristic (ROC) curve, or AUC. The traditional approach to assess the uncertainty of the estimated AUC only considers the finite testing set as the source of variability. However, a finite training set is also a random sample and the AUC varies with varying training sets. We categorize the assessment of classifiers into three levels and provide analytical expressions for the variance of the estimated AUC at each level: (1) training treated as a fixed effect, the estimated performance generalizable only to the population of testing sets; (2) training treated as a random effect, the estimated performance generalizable to both the population of training sets and the population of testing sets; (3) training treated as a random effect, performance averaged over training sets generalizable to both the population of training sets and the population of testing sets. The two sources of variability - training and testing - in automated classifiers are analogous to readers and cases in the multi-reader multi-case (MRMC) ROC paradigm in reader studies. We show the one-to-one analogy between the automated classifiers and human readers at these three levels as well as the practical difference in estimating their performance, especially regarding variance.

Ultrasonography is a valuable technique for diagnosing breast cancer. Computer-aided tumor detection and segmentation in ultrasound images can reduce labor cost and streamline clinic workflows. In this paper, we propose a fully automatic system to detect and segment breast tumors in 2D ultrasound images. Our system, based on database-guided techniques, learns the knowledge of breast tumor appearance exemplified by expert annotations. For tumor detection, we train a classifier to discriminate between tumors and their background. For tumor segmentation, we propose a discriminative graph cut approach, where both the data fidelity and compatibility functions are learned discriminatively. The performance of the proposed algorithms is demonstrated on a large set of 347 images, achieving a mean contour-to-contour error of 3.75 pixels with about 4.33 seconds.

We have previously reported an interactive information-theoretic CADe (IT-CADe) system for the detection of masses in screening mammograms. The system operates in either traditional static mode or in interactive mode whenever the user requests a second opinion. In this study we report preliminary investigation of a new paradigm of clinical integration, guided by the user's eye-gazing and reporting patterns. An observer study was conducted in which 6 radiologists evaluated 20 mammographic cases while wearing a head-mounted eye-tracking device. For each radiologistreported location, eye-gazing data were collected. Image locations that attracted prolonged dwelling (>1000msec) but were not reported were also recorded. Fixed size regions of interest (ROIs) were extracted around all above locations and analyzed using the IT-CADe system. Preliminary analysis showed that IT-CADe correctly confirmed 100% of reported true mass locations while eliminating 12.5% of the reported false positive locations. For unreported locations that attracted long dwelling, IT-CADe identified 4/6 false negative errors (i.e., errors of decision) while overcalling 8/84 TN decisions. Finally, for missed true masses that attracted short (i.e., errors of recognition) or no dwelling at all (i.e., errors of search), IT-CADe detected 5/8 of them. These results suggest that IT-CADe customization to the user's eye-gazing and reporting pattern could potentially help delineate the various sources of diagnostic error (search, recognition, decision) for each individual user and provide targeted decision support, thus improving the human-CAD synergy.

A learning-based approach integrating the use of pixel level statistical modeling and spiculation detection is presented for the segmentation of mammographic masses with ill-defined margins and spiculations. The algorithm involves a multi-phase pixel-level classification, using a comprehensive group of regional features, to generate a pixel level mass-conditional probability map (PM). Then, mass candidate along with background clutters are extracted from the PM by integrating the prior knowledge of shape and location of masses. A multi-scale steerable ridge detection algorithm is employed to detect spiculations. Finally, all the object level findings, including mass candidate, detected spiculations, and clutters, along with the PM are integrated by graph cuts to generate the final segmentation mask. The method was tested on 54 masses (51 malignant and 3 benign), all with ill-defined margins and irregular shape or spiculations. The ground truth delineations were provided by five experienced radiologists. Area overlap ratio of 0.766 (±0.144) and 0.642 (±0.173) were obtained for segmenting the whole mass and only the margin portion, respectively. Williams index of area and contour based measurements indicated that segmentation results of the algorithm well agreed with the radiologists' delineation. Most importantly, the proposed approach is capable of including mass margin and its extension which are considered as key features for breast lesion analyses.

This paper presents methods for the detection of architectural distortion in mammograms of interval-cancer cases taken prior to the diagnosis of breast cancer, using Gabor filters, phase portrait analysis, fractal dimension (FD), and analysis of the angular spread of power in the Fourier spectrum. In the estimation of FD using the Fourier power spectrum, only the distribution of power over radial frequency is considered; the information regarding the angular spread of power is ignored. In this study, the angular spread of power in the Fourier spectrum is used to generate features for the detection of spiculated patterns related to architectural distortion. Using Gabor filters and phase portrait analysis, a total of 4224 regions of interest (ROIs) were automatically obtained from 106 prior mammograms of 56 interval-cancer cases, including 301 ROIs related to architectural distortion, and from 52 mammograms of 13 normal cases. For each ROI, the FD and measures of the angular spread of power were computed. Feature selection was performed using stepwise logistic regression. The best result achieved, in terms of the area under the receiver operating characteristic curve, is 0.75 ± 0.02 with an artificial neural network including radial basis functions. Analysis of the performance of the methods with free-response receiver operating characteristics indicated a sensitivity of 0.82 at 7.7 false positives per image.

We performed a study to compare methods for volumetric breast density estimation in digital mammography (DM) and magnetic resonance imaging (MRI) for a high-risk population of women. DM and MRI images of the unaffected breast from 32 women with recently detected abnormalities and/or previously diagnosed breast cancer (age range 31-78 yrs, mean 50.3 yrs) were retrospectively analyzed. DM images were analyzed using QuantraTM (Hologic Inc). The MRI images were analyzed using a fuzzy-C-means segmentation algorithm on the T1 map. Both methods were compared to Cumulus (Univ. Toronto). Volumetric breast density estimates from DM and MRI are highly correlated (r=0.90, p≤0.001). The correlation between the volumetric and the area-based density measures is lower and depends on the training background of the Cumulus software user (r=0.73-84, p≤0.001). In terms of absolute values, MRI provides the lowest volumetric estimates (mean=14.63%), followed by the DM volumetric (mean=22.72%) and area-based measures (mean=29.35%). The MRI estimates of the fibroglandular volume are statistically significantly lower than the DM estimates for women with very low-density breasts (p≤0.001). We attribute these differences to potential partial volume effects in MRI and differences in the computational aspects of the image analysis methods in MRI and DM. The good correlation between the volumetric and the area-based measures, shown to correlate with breast cancer risk, suggests that both DM and MRI volumetric breast density measures can aid in breast cancer risk assessment. Further work is underway to fully-investigate the association between volumetric breast density measures and breast cancer risk.

We are investigating the feasibility of improving breast cancer risk prediction by computerized mammographic parenchymal pattern (MPP) analysis. A case-control study was conducted to investigate the association of the MPP measures with breast cancer risk. The case group included 168 contralateral CC-view mammograms of breast cancer patients dated at least one year prior to cancer diagnosis, and the control group included 522 CC-view mammograms from one breast of normal subjects. We extracted and compared four types of statistical texture feature spaces that included run length statistics and region size statistics (RLS/RSS) features, spatial gray level dependence (SGLD) features, gray level difference statistics (GLDS) features, and the feature space combining these three types of texture features. A linear discriminant analysis (LDA) classifier with stepwise feature selection was trained and tested with leave-one-case-out resampling to evaluate whether the breast parenchyma of future cancer patients could be distinguished from those of normal subjects in each feature space. The areas under ROC curves (Az) were 0.71, 0.72, 0.71 and 0.76 for the four feature spaces, respectively. The Az obtained from the combined feature space was significantly (p<0.05) higher than those from the individual feature spaces. Odd ratios (OR) were used to assess the association between breast cancer risk and four categories of MPP measures: <0.1 (C1), 0.1-0.15 (C2), 0.15-0.2 (C3), and >0.2 (C4) while patient age was treated as a confounding factor. The adjusted ORs of breast cancer for C2, C3 and C4 were 3.23, 7.77 and 25.43, respectively. The preliminary result indicated that our proposed computerized MPP measures were strongly associated with breast cancer risk.

A large number of false positives (FPs) generated by computer-aided detection schemes is likely to distract radiologists' attention and decreases their interpretation efficiency. Therefore, it is desirable to reduce FPs as many as possible to increase the detection specificity while maintaining the high detection sensitivity. In this paper, several features are extracted from the projected images of each initial polyp candidate to differentiate FPs from true positives. These features demonstrate the potential to exclude different types of FPs, like haustral folds, rectal tubes and residue stool by an evaluation using a database of 325 patient studies (from two different institutions) which includes 556 scans at supine and/or prone positions with 347 polyps and masses sized from 5 to 60 mm. For comparison purpose, several wellestablished features are used to generate a baseline reference. At the by-polyp detection sensitivity level of 96% (no loss of detection sensitivity), the number of FPs per scan is 7.8 by the baseline and 3.75 if the new projection features are added, which is a reduction of 51.9% FPs from the baseline.

Partial volume effect and inhomogeneity are two major causes of artifacts in electronic cleansing (EC) for non-cathartic CT colonography (CTC). Our purpose was to develop a novel method of EC for non-cathartic dual-energy CTC (DECTC) using a subvoxel multi-spectral material classifier and a regional material decomposition method for differentiation of residual fecal materials from colonic soft-tissue structures. In this study, an anthropomorphic colon phantom, which was filled with a mixture of aqueous fiber (psyllium), ground foodstuff (cereal), and non-ionic iodinated agent (Omnipaque iohexol, GE Healthcare, Milwaukee, WI), was scanned by a dual-energy CT scanner (SOMATON, Siemens) with two photon energies: 80 kVp and 140 kVp. The DE-CTC images were subjected to a dual-energy EC (DE-EC) scheme, in which a multi-spectral material classifier was used to compute the fraction of each material within one voxel by an expectation-maximization (EM) algorithm. This was followed by a regional material segmentation method for identifying of homogeneous sub-regions (tiles) as fecal materials from other tissue types. The results were compared with the structural-analysis cleansing (SA-EC) method based upon the CTC images of native phantom without fillings. The mean cleansing ratio of the DE-EC scheme was 96.57±1.21% compared to 76.3±5.56% of the SA-EC scheme. The soft-tissue preservation ratio of the DE-EC scheme was 97.05%±0.64% compared to 99.25±0.77% of the SA-EC scheme.

Predicting the malignancy of colonic polyps is a difficult problem and in general requires an invasive polypectomy procedure. We present a less-invasive and automated method to predict the histology of colonic polyps under computed tomographic colonography (CTC) using the content-based image retrieval (CBIR) paradigm. For the purpose of simplification, polyps annotated as hyperplastic or "other benign" were classified as benign polyps (BP) and the rest (adenomas and cancers) were classified as malignant polyps (MP). The CBIR uses numerical feature vectors generated from our CTC computer aided detection (CTC-CAD) system to describe the polyps. These features relate to physical and visual characteristics of the polyp. A representative database of CTC-CAD polyp images is created. Query polyps are matched with those in the database and the results are ranked based on the similarity to the query. Polyps with a majority of representative MPs in their result set are predicted to be malignant and similarly those with a majority of BPs in their results are benign. For evaluation, the system is compared to the typical optical colonoscopy (OC) size based classification. Using receiver operating curve (ROC) analysis, we show our system is sufficiently better than the OC size method.

Computed tomographic colonography (CTC) combined with a computer aided detection system has the potential for improving colonic polyp detection and increasing the use of CTC for colon cancer screening. In the clinical use of CTC, a true colonic polyp will be confirmed with high confidence if a radiologist can find it on both the supine and prone scans. To assist radiologists in CTC reading, we propose a new method for matching polyp findings on the supine and prone scans. The method performs a colon registration using four automatically identified anatomical salient points and correlation optimized warping (COW) of colon centerline features. We first exclude false positive detections using prediction information from a support vector machine (SVM) classifier committee to reduce initial false positive pairs. Then each remaining CAD detection is mapped to the other scan using COW technique applied to the distance along the centerline in each colon. In the last step, a new SVM classifier is applied to the candidate pair dataset to find true polyp pairs between supine and prone scans. Experimental results show that our method can improve the sensitivity to 0.87 at 4 false positive pairs per patient compared with 0.72 for a competing method that uses the normalized distance along the colon centerline (p<0.01).

For pharmacokinetic (PK) analysis of Dynamic Contrast Enhanced (DCE) MRI the arterial input function needs to be estimated. Previously, we demonstrated that PK parameters have a significant better discriminative performance when per patient reference tissue was used, but required manual annotation of reference tissue. In this study we propose a fully automated reference tissue segmentation method that tackles this limitation. The method was tested with our Computer Aided Diagnosis (CADx) system to study the effect on the discriminating performance for differentiating prostate cancer from benign areas in the peripheral zone (PZ). The proposed method automatically segments normal PZ tissue from DCE derived data. First, the bladder is segmented in the start-to-enhance map using the Otsu histogram threshold selection method. Second, the prostate is detected by applying a multi-scale Hessian filter to the relative enhancement map. Third, normal PZ tissue was segmented by threshold and morphological operators. The resulting segmentation was used as reference tissue to estimate the PK parameters. In 39 consecutive patients carcinoma, benign and normal tissue were annotated on MR images by a radiologist and a researcher using whole mount step-section histopathology as reference. PK parameters were computed for each ROI. Features were extracted from the set of ROIs using percentiles to train a support vector machine that was used as classifier. Prospective performance was estimated by means of leave-one-patient-out cross validation. A bootstrap resampling approach with 10,000 iterations was used for estimating the bootstrap mean AUCs and 95% confidence intervals. In total 42 malignant, 29 benign and 37 normal regions were annotated. For all patients, normal PZ was successfully segmented. The diagnostic accuracy obtained for differentiating malignant from benign lesions using a conventional general patient plasma profile showed an accuracy of 0.64 (0.53-0.74). Using the automated per-patient calibration method the diagnostic performance improved significantly to 0.76 (0.67-0.86, p=0.017) , whereas the manual per-patient calibration showed a diagnostic performance of 0.79 (0.70-0.89, p=0.01). In conclusion, the results show that an automated per-patient reference tissue PK model is feasible. A significantly better discriminating performance compared to the conventional general calibration was obtained and the diagnostic accuracy is similar to using manual per-patient calibration.

Pathological myopia is the seventh leading cause of blindness. We introduce a framework based on PAMELA (PAthological Myopia dEtection through peripapilLary Atrophy) for the detection of pathological myopia from fundus images. The framework consists of a pre-processing stage which extracts a region of interest centered on the optic disc. Subsequently, three analysis modules focus on detecting specific visual indicators. The optic disc tilt ratio module gives a measure of the axial elongation of the eye through inference from the deformation of the optic disc. In the texturebased ROI assessment module, contextual knowledge is used to demarcate the ROI into four distinct, clinically-relevant zones in which information from an entropy transform of the ROI is analyzed and metrics generated. In particular, the preferential appearance of peripapillary atrophy (PPA) in the temporal zone compared to the nasal zone is utilized by calculating ratios of the metrics. The PPA detection module obtains an outer boundary through a level-set method, and subtracts this region against the optic disc boundary. Temporal and nasal zones are obtained from the remnants to generate associated hue and color values. The outputs of the three modules are used as in a SVM model to determine the presence of pathological myopia in a retinal fundus image. Using images from the Singapore Eye Research Institute, the proposed framework reported an optimized accuracy of 90% and a sensitivity and specificity of 0.85 and 0.95 respectively, indicating promise for the use of the proposed system as a screening tool for pathological myopia.

Diabetic retinopathy (DR) is one of the leading causes of blindness among adult Americans. Automatic methods for detection of the disease have been developed in recent years, most of them addressing the segmentation of bright and red lesions. In this paper we present an automatic DR screening system that does approach the problem through the segmentation of features. The algorithm determines non-diseased retinal images from those with pathology based on textural features obtained using multiscale Amplitude Modulation-Frequency Modulation (AM-FM) decompositions. The decomposition is represented as features that are the inputs to a classifier. The algorithm achieves 0.88 area under the ROC curve (AROC) for a set of 280 images from the MESSIDOR database. The algorithm is then used to analyze the effects of image compression and degradation, which will be present in most actual clinical or screening environments. Results show that the algorithm is insensitive to illumination variations, but high rates of compression and large blurring effects degrade its performance.

A lower ratio between the width of the arteries and veins (Arteriolar-to-Venular diameter Ratio, AVR) on the retina, is well established to be predictive of stroke and other cardiovascular events in adults, as well as an increased risk of retinopathy of prematurity in premature infants. This work presents an automatic method that detects the location of the optic disc, determines the appropriate region of interest (ROI), classifies the vessels in the ROI into arteries and veins, measures their widths and calculates the AVR. After vessel segmentation and vessel width determination the optic disc is located and the system eliminates all vessels outside the AVR measurement ROI. The remaining vessels are thinned, vessel crossing and bifurcation points are removed leaving a set of vessel segments containing centerline pixels. Features are extracted from each centerline pixel that are used to assign them a soft label indicating the likelihood the pixel is part of a vein. As all centerline pixels in a connected segment should be the same type, the median soft label is assigned to each centerline pixel in the segment. Next artery vein pairs are matched using an iterative algorithm and the widths of the vessels is used to calculate the AVR. We train and test the algorithm using a set of 25 high resolution digital color fundus photographs a reference standard that indicates for the major vessels in the images whether they are an artery or a vein. We compared the AVR values produced by our system with those determined using a computer assisted method in 15 high resolution digital color fundus photographs and obtained a correlation coefficient of 0.881.

Abnormalities of retinal vasculatures can indicate health conditions in the body, such as the high blood pressure and diabetes. Providing automatically determined width ratio of arteries and veins (A/V ratio) on retinal fundus images may help physicians in the diagnosis of hypertensive retinopathy, which may cause blindness. The purpose of this study was to detect major retinal vessels and classify them into arteries and veins for the determination of A/V ratio. Images used in this study were obtained from DRIVE database, which consists of 20 cases each for training and testing vessel detection algorithms. Starting with the reference standard of vasculature segmentation provided in the database, major arteries and veins each in the upper and lower temporal regions were manually selected for establishing the gold standard. We applied the black top-hat transformation and double-ring filter to detect retinal blood vessels. From the extracted vessels, large vessels extending from the optic disc to temporal regions were selected as target vessels for calculation of A/V ratio. Image features were extracted from the vessel segments from quarter-disc to one disc diameter from the edge of optic discs. The target segments in the training cases were classified into arteries and veins by using the linear discriminant analysis, and the selected parameters were applied to those in the test cases. Out of 40 pairs, 30 pairs (75%) of arteries and veins in the 20 test cases were correctly classified. The result can be used for the automated calculation of A/V ratio.

The mapping of genotype to the phenotype of age-related macular degeneration (AMD) is expected to improve the diagnosis and treatment of the disease in a near future. In this study, we focused on the first step to discover this mapping: we identified visual patterns related to AMD which seem to be controlled by genetic factors, without explicitly relating them to the genes. For this purpose, we used a dataset of eye fundus photographs from 74 twin pairs, either monozygotic twins, who have the same genotype, or dizygotic twins, whose genes responsible for AMD are less likely to be identical. If we are able to differentiate monozygotic twins from dizygotic twins, based on a given visual pattern, then this pattern is likely to be controlled by genetic factors. The main visible consequence of AMD is the apparition of drusen between the retinal pigment epithelium and Bruch's membrane. We developed two automated drusen detectors based on the wavelet transform: a shape-based detector for hard drusen, and a texture- and color- based detector for soft drusen. Forty visual features were evaluated at the location of the automatically detected drusen. These features characterize the texture, the shape, the color, the spatial distribution, or the amount of drusen. A distance measure between twin pairs was defined for each visual feature; a smaller distance should be measured between monozygotic twins for visual features controlled by genetic factors. The predictions of several visual features (75.7% accuracy) are comparable or better than the predictions of human experts.

In this paper we present a system for fast and accurate detection of anatomical structures (calipers) in M-mode images. The task is challenging because of dramatic variations in their appearances. We propose to solve the problem in a progressive manner, which ensures both robustness and efficiency. It first obtains rough caliper localization using the intensity profile image. Then run a constrained search for accurate caliper positions. Markov Random Field (MRF) and warping image detectors are used for jointly considering appearance information and the geometric relationship between calipers. Extensive experiments show that our system achieves more accurate results and uses less time in comparison with previously reported work.

This work presents a system for automatic coronary calcium scoring and cardiovascular risk stratification in thoracic CT scans. Data was collected from a Dutch-Belgian lung cancer screening trial. In 121 low-dose, non-ECG synchronized, non-contrast enhanced thoracic CT scans an expert scored coronary calcifications manually. A key element of the proposed algorithm is that the approximate position of the coronary arteries was inferred with a probabilistic coronary calcium atlas. This atlas was created with atlas-based segmentation from 51 scans and their manually identified calcifications, and was registered to each unseen test scan. In the test scans all objects with density above 130 HU were considered candidates that could represent coronary calcifications. A statistical pattern recognition system was designed to classify these candidates using features that encode their spatial position relative to the inferred position of the coronaries obtained from the atlas registration. In addition, size and texture features were computed for all candidates. Two consecutive classifiers were used to label each candidate. The system was trained with 35 and tested with another 35 scans. The detected calcifications were quantified and cardiovascular risk was determined for each subject. The system detected 71% of coronary calcifications with an average of 0.9 false positive objects per scan. Cardiovascular risk category was correctly assigned to 29 out of 35 subjects (83%). Five scans (14%) were one category off, and only one scan (3%) was two categories off. We conclude that automatic assessment of the cardiovascular risk from low-dose, non-ECG synchronized thoracic CT scans appears feasible.

We developed a hybrid CPU-GPU framework enabling semi-automated segmentation of abdominal aortic aneurysm (AAA) on Computed Tomography Angiography (CTA) examinations. AAA maximal diameter (D-max) and volume measurements and their progression between 2 examinations can be generated by this software improving patient followup. In order to improve the workflow efficiency some segmentation tasks were implemented and executed on the graphics processing unit (GPU). A GPU based algorithm is used to automatically segment the lumen of the aneurysm within short computing time. In a second step, the user interacted with the software to validate the boundaries of the intra-luminal thrombus (ILT) on GPU-based curved image reformation. Automatic computation of D-max and volume were performed on the 3D AAA model. Clinical validation was conducted on 34 patients having 2 consecutive MDCT examinations within a minimum interval of 6 months. The AAA segmentation was performed twice by a experienced radiologist (reference standard) and once by 3 unsupervised technologists on all 68 MDCT. The ICC for intra-observer reproducibility was 0.992 (≥0.987) for D-max and 0.998 (≥0.994) for volume measurement. The ICC for inter-observer reproducibility was 0.985 (0.977-0.90) for D-max and 0.998 (0.996- 0.999) for volume measurement. Semi-automated AAA segmentation for volume follow-up was more than twice as sensitive than D-max follow-up, while providing an equivalent reproducibility.

Coronary CT angiography (cCTA) has been reported to be an effective means for diagnosis of coronary artery disease. We are investigating the feasibility of developing a computer-aided detection (CADe) system to assist radiologists in detection of non-calcified plaques in coronary arteries in ECG-gated cCTA scans. In this study, we developed a prototype vessel segmentation and tracking method to extract the coronary arterial trees which will define the search space for plaque detection. Vascular structures are first enhanced by 3D multi-scale filtering and analysis of the eigenvalues of Hessian matrices using a vessel enhancement response function specifically designed for coronary arteries. The enhanced vascular structures are then segmented by an EM estimation method. The segmented coronary arteries are tracked using a 3D dynamic balloon tracking (DBT) method. For this preliminary study, two starting seed points were manually identified at the origins of the left and right coronary artery (LCA and RCA). The DBT method automatically moves a sphere along the vessel whose diameter is adjusted dynamically based on the local vessel size, tracks the vessels, and identifies its branches to generate the left and right coronary arterial trees. The algorithm was applied to 20 cCTA scans that contained various degrees of coronary artery diseases. To evaluate the performance of vessel segmentation and tracking, the rendered volume of coronary arteries tracked by our algorithm was displayed on a PC, placed next to a GE Advantage workstation on which the coronary arterial trees tracked by the GE software and the original cCTA scan were displayed. Two experienced thoracic radiologists visually examined the coronary arteries on the cCTA scan and the segmented vessels to count untracked false-negative (FN) segments and false positives (FPs). The comparison was made by radiologists' visual judgment because the digital files for the segmented vessels were not accessible on the commercial system. A total of 19 and 38 artery segments were identified to be FNs, and 23 FPs and 20 FPs were found in the coronary trees tracked by our algorithm and the GE software, respectively. The preliminary results demonstrated the feasibility of our approach.

Despite general acceptance that a healthy lifestyle and the treatment of risk factors can prevent the development of cardiovascular diseases (CVD), CVD are the most common cause of death in Europe and the United States. It has been shown that abdominal aortic calcifications (AAC) correlate strongly with coronary artery calcifications. Hence an early detection of aortic calcified plaques helps to predict the risk of related coronary diseases. Also since two thirds of the adverse events have no prior symptoms, possibilities to screen for risk in low cost imaging are important. To this end the Morphological Atherosclerotic Calcification Distribution (MACD) index was developed. In the following several potential severity scores relating to the geometrical outline of the calcified deposits in the lumbar aortic region are introduced. Their individual as well as their combined predictive power is examined and a combined marker, MACD, is constructed. This is done using a Cox regression analysis, also known as survival analysis. Furthermore we show how a Cox regression yields MACD to be the most efficient marker. We also demonstrate that MACD has a larger individual predictive power than any of the other individual imaging markers described. Finally we present that the MACD index predicts cardiovascular death with a hazard ratio of approximately four.

Computer assisted detection (CAD) of lymph node metastases may help reduce reading time and improve interpretation of the large amount of image data in an MR-lymphography exam. We compared the influence of using different segmentation methods on the performance of a CAD system for classification of normal and metastasized lymph nodes. Our database consisted of T1 and T2*-weighted pelvic MR images of 603 lymph nodes, enhanced by USPIO contrast medium. For each lymph node, one seed point was manually defined Three automated segmentation methods were compared: 1. Confidence Connected segmentation, extended with automated Bandwidth Factor selection; 2. Conventional Graph Cut segmentation; 3. Pseudo-segmentation by selecting a sphere around the seed point. All lymph nodes were also manually segmented by a radiologist. The resulting segmentations were used to calculate 2 features (mean T1 and T2* signal intensity). Linear discriminant analysis was used for classification. The diagnostic accuracy (AUC at ROC-analysis) was: 0.95 (Confidence- Connected); 0.95 (Graph-Cut); 0.85 (spheres); and 0.95 (manual segmentations). The CAD performance of both the Confidence Connected and Graph Cut methods was as good as the manual segmentation. The substantially lower performance of the sphere segmentations demonstrates the need for accurate segmentations, even in USPIOenhanced images.

Automatic liver segmentation on CT images is challenging because the liver often abuts other organs of a similar density. Our purpose was to develop an accurate automated liver segmentation scheme for measuring liver volumes. We developed an automated volumetry scheme for the liver in CT based on a 5 step schema. First, an anisotropic smoothing filter was applied to portal-venous phase CT images to remove noise while preserving the liver structure, followed by an edge enhancer to enhance the liver boundary. By using the boundary-enhanced image as a speed function, a fastmarching algorithm generated an initial surface that roughly estimated the liver shape. A geodesic-active-contour segmentation algorithm coupled with level-set contour-evolution refined the initial surface so as to more precisely fit the liver boundary. The liver volume was calculated based on the refined liver surface. Hepatic CT scans of eighteen prospective liver donors were obtained under a liver transplant protocol with a multi-detector CT system. Automated liver volumes obtained were compared with those manually traced by a radiologist, used as "gold standard." The mean liver volume obtained with our scheme was 1,520 cc, whereas the mean manual volume was 1,486 cc, with the mean absolute difference of 104 cc (7.0%). CT liver volumetrics based on an automated scheme agreed excellently with "goldstandard" manual volumetrics (intra-class correlation coefficient was 0.95) with no statistically significant difference (p(F≤f)=0.32), and required substantially less completion time. Our automated scheme provides an efficient and accurate way of measuring liver volumes.

We present a novel non-invasive automatic method for the classification and grading of liver fibrosis from fMRI maps based on hepatic hemodynamic changes. This method automatically creates a model for liver fibrosis grading based on training datasets. Our supervised learning method evaluates hepatic hemodynamics from an anatomical MRI image and three T2*-W fMRI signal intensity time-course scans acquired during the breathing of air, air-carbon dioxide, and carbogen. It constructs a statistical model of liver fibrosis from these fMRI scans using a binary-based one-against-all multi class Support Vector Machine (SVM) classifier. We evaluated the resulting classification model with the leave-one out technique and compared it to both full multi-class SVM and K-Nearest Neighbor (KNN) classifications. Our experimental study analyzed 57 slice sets from 13 mice, and yielded a 98.2% separation accuracy between healthy and low grade fibrotic subjects, and an overall accuracy of 84.2% for fibrosis grading. These results are better than the existing image-based methods which can only discriminate between healthy and high grade fibrosis subjects. With appropriate extensions, our method may be used for non-invasive classification and progression monitoring of liver fibrosis in human patients instead of more invasive approaches, such as biopsy or contrast-enhanced imaging.

Central-chest lymph nodes play a vital role in lung-cancer staging. The three-dimensional (3D) definition of lymph nodes from multidetector computed-tomography (MDCT) images, however, remains an open problem. This is because of the limitations in the MDCT imaging of soft-tissue structures and the complicated phenomena that influence the appearance of a lymph node in an MDCT image. In the past, we have made significant efforts toward developing (1) live-wire-based segmentation methods for defining 2D and 3D chest structures and (2) a computer-based system for automatic definition and interactive visualization of the Mountain central-chest lymph-node stations. Based on these works, we propose new single-click and single-section live-wire methods for segmenting central-chest lymph nodes. The single-click live wire only requires the user to select an object pixel on one 2D MDCT section and is designed for typical lymph nodes. The single-section live wire requires the user to process one selected 2D section using standard 2D live wire, but it is more robust. We applied these methods to the segmentation of 20 lymph nodes from two human MDCT chest scans (10 per scan) drawn from our ground-truth database. The single-click live wire segmented 75% of the selected nodes successfully and reproducibly, while the success rate for the single-section live wire was 85%. We are able to segment the remaining nodes, using our previously derived (but more interaction intense) 2D live-wire method incorporated in our lymph-node analysis system. Both proposed methods are reliable and applicable to a wide range of pulmonary lymph nodes.

Many malignant processes cause abdominal lymphadenopathy, and computed tomography (CT) has become the primary modality for its detection. A lymph node is considered enlarged (swollen) if it is more than 1 centimeter in diameter. Which lymph nodes are swollen depends on the type of disease and the body parts involved. Identifying their locations is very important to determine the possible cause. In the current clinical workflow, the detection and diagnosis of enlarged lymph nodes is usually performed manually by examining all slices of CT images, which can be error-prone and time consuming. 3D blob enhancement filter is a usual way for computer-aided node detection. We proposed a new 3D blob detector for automatic lymph node detection in contrast-enhanced abdominal CT images. Since lymph nodes are usually next to blood vessels, abdominal blood vessels were first segmented as a reference to set the search region for lymph nodes. Then a new detection response measure, blobness, is defined based on eigenvalues of the Hessian matrix and the object scale in our new blob detector. Voxels with higher blobness were clustered as lymph node candidates. Finally some prior anatomical knowledge was utilized for false positive reduction. We applied our method to 5 patients and compared the results with the performance of the original blobness definition. Both methods achieved sensitivity of 83.3% but the false positive rates per patient were 14 and 26 for our method and the original method, respectively. Our results indicated that computer-aided lymph node detection with this new blob detector may yield a high sensitivity and a relatively low FP rate in abdominal CT.

Hypodense metastases are not always completely distinguishable from benign cysts in the liver using conventional Computed Tomography (CT) imaging, since the two lesion types present with overlapping intensity distributions due to similar composition as well as other factors including beam hardening and patient motion. This problem is extremely challenging for small lesions with diameter less than 1 cm. To accurately characterize such lesions, multiple follow-up CT scans or additional Positron Emission Tomography or Magnetic Resonance Imaging exam are often conducted, and in some cases a biopsy may be required after the initial CT finding. Gemstone Spectral Imaging (GSI) with fast kVp switching enables projection-based material decomposition, offering the opportunity to discriminate tissue types based on their energy-sensitive material attenuation and density. GSI can be used to obtain monochromatic images where beam hardening is reduced or eliminated and the images come inherently pre-registered due to the fast kVp switching acquisition. We present a supervised learning method for discriminating between cysts and hypodense liver metastases using these monochromatic images. Intensity-based statistical features extracted from voxels inside the lesion are used to train optimal linear and nonlinear classifiers. Our algorithm only requires a region of interest within the lesion in order to compute relevant features and perform classification, thus eliminating the need for an accurate segmentation of the lesion. We report classifier performance using M-fold cross-validation on a large lesion database with radiologist-provided lesion location and labels as the reference standard. Our results demonstrate that (a) classification using a single projection-based spectral CT image, i.e., a monochromatic image at a specified keV, outperforms classification using an image-based dual energy CT pair, i.e., low and high kVp images derived from the same fast kVp acquisition and (b) classification using monochromatic images can achieve very high accuracy in separating benign liver cysts and metastases, especially for small lesions.

Experienced radiologists are in short supply, and are sometimes called upon to read many images in a short amount of time. This leaves them with a limited amount of time to read images, and can lead to fatigue and stress which can be sources of error, as they overlook subtle abnormalities that they otherwise might not miss. Another factor in error rates is called satisfaction of search, where a radiologist misses a second (typically subtle) abnormality after finding the first. These types of errors are due primarily to a lack of attention to an important region of the image during the search. In this paper we discuss the use of eye tracker technology, in combination with image analysis and machine learning techniques, to learn what types of features catch the eye experienced radiologists when reading chest x-rays for diagnostic purposes, and to then use that information to produce saliency maps that predict what regions of each image might be most interesting to radiologists. We found that, out of 13 popular features types that are widely extracted to characterize images, 4 are particularly useful for this task: (1) Localized Edge Orientation Histograms (2) Haar Wavelets, (3) Gabor Filters, and (4) Steerable Filters.

This study describes a system for interactive annotation of thoracic CT scans. Lung volumes in these scans are segmented and subdivided into roughly spherical volumes of interest (VOIs) with homogeneous texture using a clustering procedure. For each 3D VOI, 72 features are calculated. The observer inspects the scan to determine which textures are present and annotates, with mouse clicks, several VOIs of each texture. Based on these annotations, a k-nearest-neighbor classifier is trained, which classifies all remaining VOIs in the scan. The algorithm then presents a slice with suggested annotations to the user, in which the user can correct mistakes. The classifier is retrained, taking into account these new annotations, and the user is presented another slice for correction. This process continues until at least 50% of all lung voxels in the scan have been classified. The remaining VOIs are classified automatically. In this way, the entire lung volume is annotated. The system has been applied to scans of patients with usual and non-specific interstitial pneumonia. The results of interactive annotation are compared to a setup in which the user annotates all predefined VOIs manually. The interactive system is 3.7 times as fast as complete manual annotation of VOIs and differences between the methods are similar to interobserver variability. This is a first step towards precise volumetric quantitation of texture patterns in thoracic CT in clinical research and in clinical practice.

The computer aided diagnosis (CAD) of abnormalities on chest radiographs is difficult due to the presence of overlapping normal anatomy. Suppression of the normal anatomy is expected to improve performance of a CAD system, but such a method has not yet been applied to the computer detection of interstitial abnormalities such as occur in tuberculosis (TB). The aim of this research is to evaluate the effect of rib suppression on a CAD system for TB. Profiles of pixel intensities sampled perpendicular to segmented ribs were used to create a local PCA-based shape model of the rib. The model was normalized to the local background intensity and corrected for gradients perpendicular to the rib. Subsequently rib suppressed images were created by subtracting the models for each rib from the original image. The effect of rib suppression was evaluated using a CAD system for TB detection. Small square image patches were sampled randomly from 15 normal and 35 TB-affected images containing textural abnormalities. Abnormalities were outlined by a radiologist and were given a subtlety rating from 1 to 5. Features based on moments of intensity distributions of Gaussian derivative filtered images were extracted. A supervised learning approach was used to discriminate between normal and diseased image patches. The use of rib suppressed images increased the overall performance of the system, as measured by the area under the receiver operator characteristic (ROC) curve, from 0.75 to 0.78. For the more subtly rated patches (rated 1-3) the performance increased from 0.62 to 0.70.

A novel approach to feature optimization for classification of lung carcinoma using tissue images is presented. The methodology uses a combination of three characteristics of computational features: F-measure, which is a representation of each feature towards classification, inter-correlation between features and pathology based information. The metadata provided from pathological parameters is used for mapping between computational features and biological information. Multiple regression analysis maps each category of features based on how pathology information is correlated with the size and location of cancer. Relatively the computational features represented the tumor size better than the location of the cancer. Based on the three criteria associated with the features, three sets of feature subsets with individual validation are evaluated to select the optimum feature subset. Based on the results from the three stages, the knowledgebase produces the best subset of features. An improvement of 5.5% was observed for normal Vs all abnormal cases with A_z value of 0.731 and 74/114 correctly classified. The best A_z value of 0.804 with 66/84 correct classification and improvement of 21.6% was observed for normal Vs adenocarcinoma.

Topological texture features were compared in their ability to classify morphological patterns known as 'honeycombing' that are considered indicative for the presence of fibrotic interstitial lung diseases in high-resolution computed tomography (HRCT) images. For 14 patients with known occurrence of honey-combing, a stack of 70 axial, lung kernel reconstructed images were acquired from HRCT chest exams. A set of 241 regions of interest of both healthy and pathological (89) lung tissue were identified by an experienced radiologist. Texture features were extracted using six properties calculated from gray-level co-occurrence matrices (GLCM), Minkowski Dimensions (MDs), and three Minkowski Functionals (MFs, e.g. MF.euler). A k-nearest-neighbor (k-NN) classifier and a Multilayer Radial Basis Functions Network (RBFN) were optimized in a 10-fold cross-validation for each texture vector, and the classification accuracy was calculated on independent test sets as a quantitative measure of automated tissue characterization. A Wilcoxon signed-rank test was used to compare two accuracy distributions and the significance thresholds were adjusted for multiple comparisons by the Bonferroni correction. The best classification results were obtained by the MF features, which performed significantly better than all the standard GLCM and MD features (p < 0.005) for both classifiers. The highest accuracy was found for MF.euler (97.5%, 96.6%; for the k-NN and RBFN classifier, respectively). The best standard texture features were the GLCM features 'homogeneity' (91.8%, 87.2%) and 'absolute value' (90.2%, 88.5%). The results indicate that advanced topological texture features can provide superior classification performance in computer-assisted diagnosis of interstitial lung diseases when compared to standard texture analysis methods.

Radiological bone age assessment is based on local image regions of interest (ROI), such as the epiphysis or the area of carpal bones. These are compared to a standardized reference and scores determining the skeletal maturity are calculated. For computer-aided diagnosis, automatic ROI extraction and analysis is done so far mainly by heuristic approaches. Due to high variations in the imaged biological material and differences in age, gender and ethnic origin, automatic analysis is difficult and frequently requires manual interactions. On the contrary, epiphyseal regions (eROIs) can be compared to previous cases with known age by content-based image retrieval (CBIR). This requires a sufficient number of cases with reliable positioning of the eROI centers. In this first approach to bone age assessment by CBIR, we conduct leaving-oneout experiments on 1,102 left hand radiographs and 15,428 metacarpal and phalangeal eROIs from the USC hand atlas. The similarity of the eROIs is assessed by cross-correlation of 16x16 scaled eROIs. The effects of the number of eROIs, two age computation methods as well as the number of considered CBIR references are analyzed. The best results yield an error rate of 1.16 years and a standard deviation of 0.85 years. As the appearance of the hand varies naturally by up to two years, these results clearly demonstrate the applicability of the CBIR approach for bone age estimation.

This work describes a method that can discriminate between a solid pulmonary nodule and a pulmonary vessel bifurcation point at a given candidate location on a CT scan using the method of standard moments. The algorithm starts with the estimation of a spherical window around a nodule candidate center that best captures the local shape properties of the region. Then, given this window, the standard set of moments, invariant to rotation and scale is computed over the geometric representation of the region. Finally, a feature vector composed of the moment values is classified as either a nodule or a vessel bifurcation point. The performance of this technique was evaluated on a dataset containing 276 intraparenchymal nodules and 276 selected vessel bifurcation points. The method resulted in 99% sensitivity and 80% specificity in identifying nodules, which makes this technique an efficient filter for false positives reduction. Its efficiency was further evaluated on the dataset of 656 low-dose chest CT scans. Inclusion of this filter into a design of an experimental detection system resulted in up to a 69% decrease in false positive rate in detection of intraparenchymal nodules with less than 1% loss in sensitivity.

With the advent of high-resolution CT, three-dimensional (3D) methods for nodule volumetry have been introduced, with the hope that such methods will be more accurate and consistent than currently used planar measures of size. However, the error associated with volume estimation methods still needs to be quantified. Volume estimation error is multi-faceted in the sense that there is variability associated with the patient, the software tool and the CT system. A primary goal of our current research efforts is to quantify the various sources of measurement error and, when possible, minimize their effects. In order to assess the bias of an estimate, the actual value, or "truth," must be known. In this work we investigate the reliability of micro CT to determine the "true" volume of synthetic nodules. The advantage of micro CT over other truthing methods is that it can provide both absolute volume and shape information in a single measurement. In the current study we compare micro CT volume truth to weight-density truth for spherical, elliptical, spiculated and lobulated nodules with diameters from 5 to 40 mm, and densities of -630 and +100 HU. The percent differences between micro CT and weight-density volume for -630 HU nodules range from [-21.7%, -0.6%] (mean= -11.9%) and the differences for +100 HU nodules range from [-0.9%, 3.0%] (mean=1.7%).

We have previously presented a match filtered (MF) approach for estimating lung nodule size from helical multi-detector CT (MDCT) images ^[1], in which we minimized the sum of squared differences between the simulated CT templates and the actual nodule CT images. The previous study showed the potential of this approach for reducing the bias and variance in nodule size estimation. However, minimizing SSD is not statistically optimal because the noise in 3D helical CT images is correlated. The goal of this work is to investigate the noise properties and explore several approximate descriptions of the three-dimensional (3D) noise covariance for more accurate estimates. The approximations include: variance only, noise power spectrum (NPS), axial correlation, two-dimensional (2D) in-plane correlation and fully 3D correlation. We examine the effectiveness of these second-order noise approximations by applying them to our volume estimation approach with a simulation study. Our simulations show that: the variance-based pre-whitening and axial pre-whitening perform very similar to the non-prewhitening case, with accuracy (measured in RMSE) differences within 1%; the NPS based pre-whitening performs slightly better, with a 4% decrease in RMSE; the in-plane pre-whitening and 3D fully pre-whitening perform best, with about a 10% decrease in RMSE over the non-prewhitening case. The simulation results suggest that the NPS, 2D in-plane and fully 3D prewhitening can be beneficial for lung nodule size estimation, albeit with greater computational costs in determining these noise characterizations.

The segmentation of medical images is challenging because a ground truth is often not available. Computer-Aided Detection (CAD) systems are dependent on ground truth as a means of comparison; however, in many cases the ground truth is derived from only experts' opinions. When the experts disagree, it becomes impossible to discern one ground truth. In this paper, we propose an algorithm to measure the disagreement among radiologist's delineated boundaries. The algorithm accounts for both the overlap and shape of the boundaries in determining the variability of a panel segmentation. After calculating the variability of 3788 thoracic computed tomography (CT) slices in the Lung Image Database Consortium (LIDC), we found that the radiologists have a high consensus in a majority of lung nodule segmentations. However, our algorithm identified a number of segmentations that the radiologists significantly disagreed on. Our proposed method of measuring disagreement can assist others in determining the reliability of panel segmentations. We also demonstrate that it is superior to simply using overlap, which is currently one of the most common ways of measuring segmentation agreement. The variability metric presented has applications to panel segmentations, and also has potential uses in CAD systems.

As part of a more general effort to probe the interrelated factors impacting the accuracy and precision of lung nodule size estimation, we have been conducting phantom CT studies with an anthropomorphic thoracic phantom containing a vasculature insert on which synthetic nodules were inserted or attached. The utilization of synthetic nodules with known truth regarding size and location allows for bias and variance analysis, enabled by the acquisition of repeat CT scans. Using a factorial approach to probe imaging parameters (acquisition and reconstruction) and nodule characteristics (size, density, shape, location), ten repeat scans have been collected for each protocol and nodule layout. The resulting database of CT scans is incrementally becoming available to the public via the National Biomedical Imaging Archive to facilitate the assessment of lung nodule size estimation methodologies and the development of image analysis software among other possible applications. This manuscript describes the phantom CT scan database and associated information including image acquisition and reconstruction protocols, nodule layouts and nodule truth.

This work presents a novel approach for model based segmentation of the kidney in images acquired by Computed Tomography (CT). The developed computer aided segmentation system is expected to support computer aided diagnosis and operation planning. We have developed a deformable model based approach based on local shape constraints that prevents the model from deforming into neighboring structures while allowing the global shape to adapt freely to the data. Those local constraints are derived from the anatomical structure of the kidney and the presence and appearance of neighboring organs. The adaptation process is guided by a rule-based deformation logic in order to improve the robustness of the segmentation in areas of diffuse organ boundaries. Our work flow consists of two steps: 1.) a user guided positioning and 2.) an automatic model adaptation using affine and free form deformation in order to robustly extract the kidney. In cases which show pronounced pathologies, the system also offers real time mesh editing tools for a quick refinement of the segmentation result. Evaluation results based on 30 clinical cases using CT data sets show an average dice correlation coefficient of 93% compared to the ground truth. The results are therefore in most cases comparable to manual delineation. Computation times of the automatic adaptation step are lower than 6 seconds which makes the proposed system suitable for an application in clinical practice.

Intervertebral disc herniation is a major reason for lower back pain (LBP), which is the second most common neurological ailment in the United States. Automation of herniated disc diagnosis reduces the large burden on radiologists who have to diagnose hundreds of cases each day using clinical MRI. We present a method for automatic diagnosis of lumbar disc herniation using appearance and shape features. We jointly use the intensity signal for modeling the appearance of herniated disc and the active shape model for modeling the shape of herniated disc. We utilize a Gibbs distribution for classification of discs using appearance and shape features. We use 33 clinical MRI cases of the lumbar area for training and testing both appearance and shape models. We achieve over 91% accuracy in detection of herniation in a cross-validation experiment with specificity of 91% and sensitivity of 94%.

In building robust classifiers for computer-aided detection (CAD) of lesions, selection of relevant features is of fundamental importance. Typically one is interested in determining which, of a large number of potentially redundant or noisy features, are most discriminative for classification. Searching all possible subsets of features is impractical computationally. This paper proposes a feature selection scheme combining AdaBoost with the Minimum Redundancy Maximum Relevance (MRMR) to focus on the most discriminative features. A fitness function is designed to determine the optimal number of features in a forward wrapper search. Bagging is applied to reduce the variance of the classifier and make a reliable selection. Experiments demonstrate that by selecting just 11 percent of the total features, the classifier can achieve better prediction on independent test data compared to the 70 percent of the total features selected by AdaBoost.

Computer-aided diagnosis (CAD) systems are indispensable tools for patients' healthcare in modern medicine. Nevertheless, the only fully automatic CAD system available for lumbar stenosis today is for X-ray images. Its performance is limited due to the limitations intrinsic to X-ray images. In this paper, we present a system for magnetic resonance images. It employs a machine learning classification technique to automatically recognize lumbar spine components. Features can then be extracted from these spinal components. Finally, diagnosis is done by applying a Multilayer Perceptron. This classification framework can learn the features of different spinal conditions from the training images. The trained Perceptron can then be applied to diagnose new cases for various spinal conditions. Our experimental studies based on 62 subjects indicate that the proposed system is reliable and significantly better than our older system for X-ray images.

We are developing a computer-aided detection (CAD) system for clustered microcalcifications in digital breast tomosynthesis (DBT). In this preliminary study, we investigated the approach of detecting microcalcifications in the tomosynthesized volume. The DBT volume is first enhanced by 3D multi-scale filtering and analysis of the eigenvalues of Hessian matrices with a calcification response function and signal-to-noise ratio enhancement filtering. Potential signal sites are identified in the enhanced volume and local analysis is performed to further characterize each object. A 3D dynamic clustering procedure is designed to locate potential clusters using hierarchical criteria. We collected a pilot data set of two-view DBT mammograms of 39 breasts containing microcalcification clusters (17 malignant, 22 benign) with IRB approval. A total of 74 clusters were identified by an experienced radiologist in the 78 DBT views. Our prototype CAD system achieved view-based sensitivity of 90% and 80% at an average FP rate of 7.3 and 2.0 clusters per volume, respectively. At the same levels of case-based sensitivity, the FP rates were 3.6 and 1.3 clusters per volume, respectively. For the subset of malignant clusters, the view-based detection sensitivity was 94% and 82% at an average FP rate of 6.0 and 1.5 FP clusters per volume, respectively. At the same levels of case-based sensitivity, the FP rates were 1.2 and 0.9 clusters per volume, respectively. This study demonstrated that computerized microcalcification detection in 3D is a promising approach to the development of a CAD system for DBT. Study is underway to further improve the computer-vision methods and to optimize the processing parameters using a larger data set.

We present a novel method for the detection and reconstruction in 3D of microcalcifications in digital breast tomosynthesis (DBT) image sets. From a list of microcalcification candidate regions (that is, real microcalcification points or noise points) found in each DBT projection, our method: (1) finds the set of corresponding points of a microcalcification in all the other projections; (2) locates its 3D position in the breast; (3) highlights noise points; and (4) identifies the failure of microcalcification detection in one or more projections, in which case the method predicts the image locations of the microcalcification in the images in which they are missed. From the geometry of the DBT acquisition system, an "epipolar curve" is derived for the 2D positions a microcalcification in each projection generated at different angular positions. Each epipolar curve represents a single microcalcification point in the breast. By examining the n projections of m microcalcifications in DBT, one expects ideally m epipolar curves each comprising n points. Since each microcalcification point is at a different 3D position, each epipolar curve will be at a different position in the same 2D coordinate system. By plotting all the microcalcification candidates in the same 2D plane simultaneously, one can easily extract a representation of the number of microcalcification points in the breast (number of epipolar curves) and their 3D positions, the noise points detected (isolated points not forming any epipolar curve) and microcalcification points missed in some projections (epipolar curves with less than n points).

We are developing a computer-aided detection (CAD) system to assist radiologists in detecting microcalcification clusters in digital breast tomosynthesis (DBT). The purpose of this study is to investigate the feasibility of a 2D approach using the projection-view (PV) images as input. In the first stage, automated detection of the microcalcification clusters on the PVs is performed. In the second stage, the detected cluster candidates or the individual microcalcifications on the PVs are back-projected to the 3D volume. The true clusters or microcalcifications will therefore converge at their focal planes and ideally will result in higher cluster or microcalcification scores than the FPs. In the final step an analysis of the back-projected cluster or microcalcification candidates is performed to differentiate the true and false clusters. In this pilot study, a limited data set of 39 cases with biopsy proven microcalcification clusters (17 malignant, 22 benign) was used. The DBT scans were obtained in both CC and MLO views using a GE GEN2 prototype system which acquires 21 PVs over a 60º arc in 3º increments. In the 78 DBT volumes, a total of 74 clusters (33 malignant clusters in 34 breasts and 41 benign clusters in 44 breasts) were identified by an experienced radiologist. The computer detected 61% (956/1554) of the clusters on the PVs from the 74 scans. After back-projection of the microcalcification candidates detected on the individual PVs and excluding the first few PVs that had higher noise in back-projection stage, 84% (62/74) of the true clusters were detected in the 3D volume. Study is underway to develop methods to reduce FPs and to compare this 2D approach with 3D or combined 2D and 3D approaches.

Haralick texture features derived from gray-level co-occurrence matrices (GLCM) were used to classify the character of suspicious breast lesions as benign or malignant on dynamic contrast-enhanced MRI studies. Lesions were identified and annotated by an experienced radiologist on 54 MRI exams of female patients where histopathological reports were available prior to this investigation. GLCMs were then extracted from these 2D regions of interest (ROI) for four principal directions (0°, 45°, 90° & 135°) and used to compute Haralick texture features. A fuzzy k-nearest neighbor (k- NN) classifier was optimized in ten-fold cross-validation for each texture feature and the classification performance was calculated on an independent test set as a function of area under the ROC curve. The lesion ROIs were characterized by texture feature vectors containing the Haralick feature values computed from each directional-GLCM; and the classifier results obtained were compared to a previously used approach where the directional-GLCMs were summed to a nondirectional GLCM which could further yield a set of texture feature values. The impact of varying the inter-pixel distance while generating the GLCMs on the classifier's performance was also investigated. Classifier's AUC was found to significantly increase when the high-dimensional texture feature vector approach was pursued, and when features derived from GLCMs generated using different inter-pixel distances were incorporated into the classification task. These results indicate that lesion character classification accuracy could be improved by retaining the texture features derived from the different directional GLCMs rather than combining these to yield a set of scalar feature values instead.

Dynamic contrast enhanced Breast MRI (DCE BMRI) has emerged as powerful tool in the diagnostic work-up of breast cancer. While DCE BMRI is very sensitive, specificity remains to be an issue. Consequently, there is a need for features that support the classification of enhancing lesions into benign and malignant lesions. Traditional features include the morphology and the texture of a lesion, as well as the kinetic parameters of the time-intensity curves, i.e., the temporal change of image intensity at a given location. The kinetic parameters include initial contrast uptake of a lesion and the type of the kinetic curve. The curve type is usually assigned to one of three classes: persistent enhancement (Type I), plateau (Type II), and washout (Type III). While these curve types show a correlation with the tumor type (benign or malignant), only a small sub-volume of the lesion is taken into consideration and the curve type will depend on the location of the ROI that was used to generate the kinetic curve. Furthermore, it has been shown that the curve type significantly depends on which MR scanner was used as well as on the scan parameters. Recently, it was shown that the heterogeneity of a given lesion with respect to spatial variation of the kinetic curve type is a clinically significant indicator for malignancy of a tumor. In this work we compare four quantitative measures for the degree of heterogeneity of the signal enhancement ratio in a tumor and evaluate their ability of predicting the dignity of a tumor. All features are shown to have an area under the ROC curve of between 0.63 and 0.78 (for a single feature).

Previous research has shown that a fuzzy C-means (FCM) approach to computerized lesion analysis has the potential to aid radiologists in the interpretation of dynamic contrast-enhanced MRI (DCE-MRI) breast exams. ^{1, 2} Our purpose in this study was to optimize the performance of the FCM approach with respect to binary (benign/malignant) breast lesion classification in DCE-MRI. We used both raw (calculated from kinetic data points) and empirically fitted³ kinetic features for this study. FCM was used to automatically select a characteristic kinetic curve (CKC) based on intensity-time point data of voxels within each lesion, using four different kinetic criteria: (1) maximum initial enhancement, (2) minimum shape index, (3) maximum washout, and (4) minimum time to peak. We extracted kinetic features from these CKCs, which were merged using linear discriminant analysis (LDA), and evaluated with receiver operating characteristic (ROC) analysis. There was comparable performance for methods 1, 2, and 4, while method 3 was inferior. Next, we modified use of the FCM method by calculating a feature vector for every voxel in each lesion and using FCM to select a characteristic feature vector (CFV) for each lesion. Using this method, we achieved performance similar to the four CKC methods. Finally, we generated lesion color maps using FCM membership matrices, which facilitated the visualization of enhancing voxels in a given lesion.

Arterial spin labeling (ASL) is one of promising non-invasive magnetic resonance (MR) imaging techniques for diagnosis of Alzheimer's disease (AD) by measuring cerebral blood flow (CBF). The aim of this study was to develop a computer-aided classification system for AD patients based on CBFs measured by the ASL technique. The average CBFs in cortical regions were determined as functional image features based on the CBF map image, which was non-linearly transformed to a Talairach brain atlas by using a free-form deformation. An artificial neural network (ANN) was trained with the CBF functional features in 10 cortical regions, and was employed for distinguishing patients with AD from control subjects. For evaluation of the method, we applied the proposed method to 20 cases including ten AD patients and ten control subjects, who were scanned a 3.0-Tesla MR unit. As a result, the area under the receiver operating characteristic curve obtained by the proposed method was 0.893 based on a leave-one-out-by-case test in identification of AD cases among 20 cases. The proposed method would be feasible for classification of patients with AD.

In this paper, we present an approach of predictive modeling of neuroanatomic structures for the detection of brain atrophy based on cross-sectional MRI image. The underlying premise of applying predictive modeling for atrophy detection is that brain atrophy is defined as significant deviation of part of the anatomy from what the remaining normal anatomy predicts for that part. The steps of predictive modeling are as follows. The central cortical surface under consideration is reconstructed from brain tissue map and Regions of Interests (ROI) on it are predicted from other reliable anatomies. The vertex pair-wise distance between the predicted vertex and the true one within the abnormal region is expected to be larger than that of the vertex in normal brain region. Change of white matter/gray matter ratio within a spherical region is used to identify the direction of vertex displacement. In this way, the severity of brain atrophy can be defined quantitatively by the displacements of those vertices. The proposed predictive modeling method has been evaluated by using both simulated atrophies and MRI images of Alzheimer's disease.

This paper introduces a general framework for spatial prior in SVM-based classification of brain images based on Laplacian regularization. Most existing methods include spatial prior by adding a feature aggregation step before the SVM classification. The problem of the aggregation step is that the individual information of each feature is lost. Our framework enables to avoid this shortcoming by including the spatial prior directly in the SVM. We demonstrate that this framework can be used to derive embedded regularization corresponding to existing methods for classification of brain images and propose an efficient way to implement them. This framework is illustrated on the classification of MR images from 55 patients with Alzheimer's disease and 82 elderly controls selected from the ADNI database. The results demonstrate that the proposed algorithm enables introducing straightforward and anatomically consistent spatial prior into the classifier.

Slowly varying temporally correlated activity fluctuations between functionally related brain areas have been identified by functional magnetic resonance imaging (fMRI) research in recent years. These low-frequency oscillations of less than 0.08 Hz appear to play a major role in various dynamic functional brain networks, such as the so-called 'default mode' network. They also have been observed as a property of symmetric cortices, and they are known to be present in the motor cortex among others. These low-frequency data are difficult to detect and quantify in fMRI. Traditionally, user-based regions of interests (ROI) or 'seed clusters' have been the primary analysis method. In this paper, we propose unsupervised clustering algorithms based on various distance measures to detect functional connectivity in resting state fMRI. The achieved results are evaluated quantitatively for different distance measures. The Euclidian metric implemented by standard unsupervised clustering approaches is compared with a non-metric topographic mapping of proximities based on the the mutual prediction error between pixel-specific signal dynamics time-series. It is shown that functional connectivity in the motor cortex of the human brain can be detected based on such model-free analysis methods for resting state fMRI.

Multiatlas based segmentation-propagation approaches have been shown to obtain accurate parcelation of brain structures. However, this approach requires a large number of manually delineated atlases, which are often not available. We propose a supervised method to build a population specific atlas database, using the publicly available Internet Brain Segmentation Repository (IBSR). The set of atlases grows iteratively as new atlases are added, so that its segmentation capability may be enhanced in the multiatlas based approach. Using a dataset of 210 MR images of elderly subjects (170 elderly control, 40 Alzheimer's disease) from the Australian Imaging, Biomarkers and Lifestyle (AIBL) study, 40 MR images were segmented to build a population specific atlas database for the purpose of multiatlas segmentation-propagation. The population specific atlases were used to segment the elderly population of 210 MR images, and were evaluated in terms of the agreement among the propagated labels. The agreement was measured by using the entropy H of the probability image produced when fused by voting rule and the partial moment μ2 of the histogram. Compared with using IBSR atlases, the population specific atlases obtained a higher agreement when dealing with images of elderly subjects.

Airway remodeling and accompanying changes in wall thickness are known to be a major symptom of chronic obstructive pulmonary disease (COPD), associated with reduced lung function in diseased individuals. Further investigation of this disease as well as monitoring of disease progression and treatment effect demand for accurate and reproducible assessment of airway wall thickness in CT datasets. With wall thicknesses in the sub-millimeter range, this task remains challenging even with today's high resolution CT datasets. To provide accurate measurements, taking partial volume effects into account is mandatory. The Full-Width-at-Half-Maximum (FWHM) method has been shown to be inappropriate for small airways^1,2 and several improved algorithms for objective quantification of airway wall thickness have been proposed.^1-8 In this paper, we describe an algorithm based on a closed form solution proposed by Weinheimer et al.⁷ We locally estimate the lung density parameter required for the closed form solution to account for possible variations of parenchyma density between different lung regions, inspiration states and contrast agent concentrations. The general accuracy of the algorithm is evaluated using basic tubular software and hardware phantoms. Furthermore, we present results on the reproducibility of the algorithm with respect to clinical CT scans, varying reconstruction kernels, and repeated acquisitions, which is crucial for longitudinal observations.

While accurate diagnosis of pure nodular ground glass opacity (PNGGO) is important in order to reduce the number of unnecessary biopsies, computer-aided diagnosis of PNGGO is less studied than other types of pulmonary nodules (e.g., solid-type nodule). Difficulty in segmentation of GGO nodules is one of technical bottleneck in the development of CAD of GGO nodules. In this study, we propose an automated volumetric segmentation method for PNGGO using a modeling of ROI histogram with a Gaussian mixture. Our proposed method segments lungs and applies noise-filtering in the pre-processing step. And then, histogram of selected ROI is modeled as a mixture of two Gaussians representing lung parenchyma and GGO tissues. The GGO nodule is then segmented by region-growing technique that employs the histogram model as a probability density function of each pixel belonging to GGO nodule, followed by the elimination of vessel-like structure around the nodules using morphological image operations. Our results using a database of 26 cases indicate that the automated segmentation method have a promising potential.

The purpose of this work was to retrospectively investigate the variation of standard indices of pulmonary emphysema from helical computed tomographic (CT) scans as related to inspiration differences over a 1 year interval and determine the strength of the relationship between these measures in a large cohort. 626 patients that had 2 scans taken at an interval of 9 months to 15 months (μ: 381 days, σ: 31 days) were selected for this work. All scans were acquired at a 1.25mm slice thickness using a low dose protocol. For each scan, the emphysema index (EI), fractal dimension (FD), mean lung density (MLD), and 15th percentile of the histogram (HIST) were computed. The absolute and relative changes for each measure were computed and the empirical 95% confidence interval was reported both in non-normalized and normalized scales. Spearman correlation coefficients are computed between the relative change in each measure and relative change in inspiration between each scan-pair, as well as between each pair-wise combination of the four measures. EI varied on a range of -10.5 to 10.5 on a non-normalized scale and -15 to 15 on a normalized scale, with FD and MLD showing slightly larger but comparable spreads, and HIST having a much larger variation. MLD was found to show the strongest correlation to inspiration change (r=0.85, p<0.001), and EI, FD, and HIST to have moderately strong correlation (r = 0.61-0.74, p<0.001). Finally, HIST showed very strong correlation to EI (r = 0.92, p<0.001), while FD showed the least strong relationship to EI (r = 0.82, p<0.001). This work shows that emphysema index and fractal dimension have the least variability overall of the commonly used measures of emphysema and that they offer the most unique quantification of emphysema relative to each other.

Imaging is a crucial clinical tool for diagnosis and assessment of pneumonia, but quantitative methods are lacking. Micro-computed tomography (micro CT), designed for lab animals, provides opportunities for non-invasive radiographic endpoints for pneumonia studies. HYPOTHESIS: In vivo micro CT scans of mice with early bacterial pneumonia can be scored quantitatively by semiautomated imaging methods, with good reproducibility and correlation with bacterial dose inoculated, pneumonia survival outcome, and radiologists' scores. METHODS: Healthy mice had intratracheal inoculation of E. coli bacteria (n=24) or saline control (n=11). In vivo micro CT scans were performed 24 hours later with microCAT II (Siemens). Two independent radiologists scored the extent of airspace abnormality, on a scale of 0 (normal) to 24 (completely abnormal). Using the Amira 5.2 software (Mercury Computer Systems), a histogram distribution of voxel counts between the Hounsfield range of -510 to 0 was created and analyzed, and a segmentation procedure was devised. RESULTS: A t-test was performed to determine whether there was a significant difference in the mean voxel value of each mouse in the three experimental groups: Saline Survivors, Pneumonia Survivors, and Pneumonia Non-survivors. It was found that the voxel count method was able to statistically tell apart the Saline Survivors from the Pneumonia Survivors, the Saline Survivors from the Pneumonia Non-survivors, but not the Pneumonia Survivors vs. Pneumonia Non-survivors. The segmentation method, however, was successfully able to distinguish the two Pneumonia groups. CONCLUSION: We have pilot-tested an evaluation of early pneumonia in mice using micro CT and a semi-automated method for lung segmentation and scoring system. Statistical analysis indicates that the system is reliable and merits further evaluation.

A coupled surface graph cut algorithm for airway wall segmentation from Computed Tomography (CT) images is presented. Using cost functions that highlight both inner and outer wall borders, the method combines the search for both borders into one graph cut. The proposed method is evaluated on 173 manually segmented images extracted from 15 different subjects and shown to give accurate results, with 37% less errors than the Full Width at Half Maximum (FWHM) algorithm and 62% less than a similar graph cut method without coupled surfaces. Common measures of airway wall thickness such as the Interior Area (IA) and Wall Area percentage (WA%) was measured by the proposed method on a total of 723 CT scans from a lung cancer screening study. These measures were significantly different for participants with Chronic Obstructive Pulmonary Disease (COPD) compared to asymptomatic participants. Furthermore, reproducibility was good as confirmed by repeat scans and the measures correlated well with the outcomes of pulmonary function tests, demonstrating the use of the algorithm as a COPD diagnostic tool. Additionally, a new measure of airway wall thickness is proposed, Normalized Wall Intensity Sum (NWIS). NWIS is shown to correlate better with lung function test values and to be more reproducible than previous measures IA, WA% and airway wall thickness at a lumen perimeter of 10 mm (PI10).

Quantification of potentially cancerous lesions from imaging modalities, most prominently from CT or PET images, plays a crucial role both in diagnosing and staging of cancer as well as in the assessment of the response of a cancer to a therapy, e.g. for lymphoma or lung cancer. For PET imaging, several quantifications which might bear great discriminating potential (e.g. total tumor burden or total tumor glycolysis) involve the segmentation of the entirety of all of the cancerous lesions. However, this particular task of segmenting the entirety of all cancerous lesions might be very tedious if it has to be done manually, in particular if the disease is scattered or metastasized and thus consists of numerous foci; this is one of the reasons why only few clinical studies on those quantifications are available. In this work, we investigate a way to aid the easy determination of the entirety of cancerous lesions in a PET image of a human. The approach is designed to detect all hot spots within a PET image and rank their probability of being a cancerous lesion. The basis of this component is a modified watershed algorithm; the ranking is performed on a combination of several, primarily morphological measures derived from the individual basins. This component is embedded in a software suite to assess response to a therapy based on PET images. As a preprocessing step, potential lesions are segmented and indicated to the user, who can select the foci which constitute the tumor and discard the false positives. This procedure substantially simplifies the segmentation of the entire tumor burden of a patient. This approach of semi-automatic hot spot detection is evaluated on 17 clinical datasets.

The purpose of this paper is to present a CAD prototype system for Crohn's disease. Crohn's disease causes inflammation or ulcers of the gastrointestinal tract. The number of patients of Crohn's disease is increasing in Japan. Symptoms of Crohn's disease include intestinal stenosis, longitudinal ulcers, and fistulae. Optical endoscope cannot pass through intestinal stenosis in some cases. We propose a new CAD system using abdominal fecal tagging CT images for efficient diagnosis of Crohn's disease. The system displays virtual unfolded (VU), virtual endoscopic, curved planar reconstruction, multi planar reconstruction, and outside views of both small and large intestines. To generate the VU views, we employ a small and large intestines extraction method followed by a simple electronic cleansing method. The intestine extraction is based on the region growing process, which uses a characteristic that tagged fluid neighbor air in the intestine. The electronic cleansing enables observation of intestinal wall under tagged fluid. We change the height of the VU views according to the perimeter of the intestine. In addition, we developed a method to enhance the longitudinal ulcer on views of the system. We enhance concave parts on the intestinal wall, which are caused by the longitudinal ulcer, based on local intensity structure analysis. We examined the small and the large intestines of eleven CT images by the proposed system. The VU views enabled efficient observation of the intestinal wall. The height change of the VU views helps finding intestinal stenosis on the VU views. The concave region enhancement made longitudinal ulcers clear on the views.

With the development of computer-aided polyp detection towards virtual colonoscopy screening, the trade-off between detection sensitivity and specificity has gained increasing attention. An optimum detection, with least number of false positives and highest true positive rate, is desirable and involves interdisciplinary knowledge, such as feature extraction, feature selection as well as machine learning. Toward that goal, various geometrical and textural features, associated with each suspicious polyp candidate, have been individually extracted and stacked together as a feature vector. However, directly inputting these high-dimensional feature vectors into a learning machine, e.g., neural network, for polyp detection may introduce redundant information due to feature correlation and induce the curse of dimensionality. In this paper, we explored an indispensable building block of computer-aided polyp detection, i.e., principal component analysis (PCA)-weighted feature selection for neural network classifier of true and false positives. The major concepts proposed in this paper include (1) the use of PCA to reduce the feature correlation, (2) the scheme of adaptively weighting each principal component (PC) by the associated eigenvalue, and (3) the selection of feature combinations via the genetic algorithm. As such, the eigenvalue is also taken as part of the characterizing feature, and the necessary number of features can be exposed to mitigate the curse of dimensionality. Learned and tested by radial basis neural network, the proposed computer-aided polyp detection has achieved 95% sensitivity at a cost of average 2.99 false positives per polyp.

Polycystic kidney disease (PKD) is a disorder characterized by the growth of numerous fluid filled cysts in the kidneys. Measuring cystic kidney volume is thus crucial to monitoring the evolution of the disease. While T2-weighted MRI delineates the organ, automatic segmentation is very difficult due to highly variable shape and image contrast. The interactive stereology methods used currently involve a compromise between segmentation accuracy and time. We have investigated semi-automated methods: active contours and a sub-voxel morphology based algorithm. Coronal T2- weighted images of 17 patients were acquired in four breath-holds using the HASTE sequence on a 1.5 Tesla MRI unit. The segmentation results were compared to ground truth kidney masks obtained as a consensus of experts. Automatic active contour algorithm yielded an average 22% ± 8.6% volume error. A recently developed method (Bridge Burner) based on thresholding and constrained morphology failed to separate PKD from the spleen, yielding 37.4% ± 8.7% volume error. Manual post-editing reduced the volume error to 3.2% ± 0.8% for active contours and 3.2% ± 0.6% for Bridge Burner. The total time (automated algorithm plus editing) was 15 min ± 5 min for active contours and 19 min ± 11 min for Bridge Burner. The average volume errors for stereology method were 5.9%, 6.2%, 5.4% for mesh size 6.6, 11, 16.5 mm. The average processing times were 17, 7, 4 min. These results show that nearly two-fold improvement in PKD segmentation accuracy over stereology technique can be achieved with a combination of active contours and postediting.

In aging societies, it is important to analyze age-related hypokinesia. A psoas major muscle has many important functional capabilities such as capacity of balance and posture control. These functions can be measured by its cross sectional area (CSA), volume, and thickness. However, these values are calculated manually in the clinical situation. The purpose of our study is to propose an automated recognition method of psoas major muscles in X-ray torso CT images. The proposed recognition process involves three steps: 1) determination of anatomical points such as the origin and insertion of the psoas major muscle, 2) generation of a shape model for the psoas major muscle, and 3) recognition of the psoas major muscles by use of the shape model. The model was built using quadratic function, and was fit to the anatomical center line of psoas major muscle. The shape model was generated using 20 CT cases and tested by 20 other CT cases. The applied database consisted of 12 male and 8 female cases from the ages of 40's to 80's. The average value of Jaccard similarity coefficient (JSC) values employed in the evaluation was 0.7. Our experimental results indicated that the proposed method was effective for a volumetric analysis and could be possible to be used for a quantitative measurement of psoas major muscles in CT images.

Co-located optical and virtual colonoscopy images have the potential to provide important clinical information during routine colonoscopy procedures. In our earlier work, we presented an optical flow based algorithm to compute egomotion from live colonoscopy video, permitting navigation and visualization of the corresponding patient anatomy. In the original algorithm, motion parameters were estimated using the traditional Least Sum of squares(LS) procedure which can be unstable in the context of optical flow vectors with large errors. In the improved algorithm, we use the Least Median of Squares (LMS) method, a robust regression method for motion parameter estimation. Using the LMS method, we iteratively analyze and converge toward the main distribution of the flow vectors, while disregarding outliers. We show through three experiments the improvement in tracking results obtained using the LMS method, in comparison to the LS estimator. The first experiment demonstrates better spatial accuracy in positioning the virtual camera in the sigmoid colon. The second and third experiments demonstrate the robustness of this estimator, resulting in longer tracked sequences: from 300 to 1310 in the ascending colon, and 410 to 1316 in the transverse colon.

We present an algorithm for the segmentation of the liver portal veins from an arterial phase CT. The developed segmentation algorithm incorporates a physiological model that states that the vasculature pattern is organized such that the whole organ is perfused using minimal mechanical energy. This model is, amongst others, applicable to the lungs, the liver, and the kidneys. The algorithm first locally detects probable candidate vessel segments in the image. The subset of these segments that generates the most probable vessel tree according the image and the physiological model is afterwards sought by a global optimization method. The algorithm has already been applied successfully to segment heavily simplified lung vessel trees from CT images. Now the general feasibility of this approach is evaluated by applying it to the segmentation of the liver portal veins from an arterial phase CT scan. This is more challenging, because the intensity difference between the vessels and the parenchyma is small. To cope with the low contrast a support vector machines approach with a robust feature vector is used to locally detect vessels. This approach has been applied to a set of five images, for which a ground truth segmentation is available. This algorithm is a first step towards an automatic segmentation of all of the liver vasculature.

This paper proposes a method for making correspondence between the supine and the prone positions of the colon in CT volumes. In CT colonography, two CT volumes in the supine and the prone positions are often taken to observe the whole colonic wall by comparing them. However, the colonic wall is soft and changes its shape when a patient changes positions. Therefore, physicians need to take the positional relations into account when comparing the two CT volumes. Calculation of the positional relations between the two positions of the colon can reduce load of physicians. A large number of haustral folds exists in the colon and the order doesn't change even when a patient change positions. Therefore, haustral folds are suitable for registering the supine and the prone positions of the colon. We also find sharply bending points of the centerline of the colon as landmarks for brief registration. The precise registration is then performed by finding positional correspondence of the haustral folds in the supine and the prone positions. In correspondence search, we first find the correspondence among long haustral folds, followed by small haustral folds. As the result of experiment using six pairs of 3D abdominal CT volumes, 65.1% of the correspondence of large haustral folds were correct, 25.6% were incorrect, and 9.3% could not be judged. On the other hand, 13.3% of the correspondence of small haustral folds were correct, 42.9% were incorrect, and 32.7% could not be judged.

Most colon CAD (computer aided detection) software products, especially commercial products, are designed for use by radiologists in a clinical environment. Therefore, those features that effectively assist radiologists in finding polyps are emphasized in those tools. However, colon CAD researchers, many of whom are engineers or computer scientists, are working with CT studies in which polyps have already been identified using CT Colonography (CTC) and/or optical colonoscopy (OC). Their goal is to utilize that data to design a computer system that will identify all true polyps with no false positive detections. Therefore, they are more concerned with how to reduce false positives and to understand the behavior of the system than how to find polyps. Thus, colon CAD researchers have different requirements for tools not found in current CAD software. We have implemented a module in 3D Slicer to assist these researchers. As with clinical colon CAD implementations, the ability to promptly locate a polyp candidate in a 2D slice image and on a 3D colon surface is essential for researchers. Our software provides this capability, and uniquely, for each polyp candidate, the prediction value from a classifier is shown next to the 3D view of the polyp candidate, as well as its CTC/OC finding. This capability makes it easier to study each false positive detection and identify its causes. We describe features in our colon CAD system that meets researchers' specific requirements. Our system uses an open source implementation of a 3D Slicer module, and the software is available to the pubic for use and for extension (http://www2.wfubmc.edu/ctc/download/).

For the early detection of prostate cancer, the analysis of the Prostate-specific antigen (PSA) in serum is currently the most popular approach. However, previous studies show that 15% of men have prostate cancer even their PSA concentrations are low. MALDI Mass Spectrometry (MS) proves to be a better technology to discover molecular tools for early cancer detection. The molecular tools or peptides are termed as biomarkers. Using MALDI MS data from prostate tissue samples, prostate cancer biomarkers can be identified by searching for molecular or molecular combination that can differentiate cancer tissue regions from normal ones. Cancer tissue regions are usually identified by pathologists after examining H&E stained histological microscopy images. Unfortunately, histopathological examination is currently done on an adjacent slice because the H&E staining process will change tissue's protein structure and it will derogate MALDI analysis if the same tissue is used, while the MALDI imaging process will destroy the tissue slice so that it is no longer available for histopathological exam. For this reason, only the most confident cancer region resulting from the histopathological examination on an adjacent slice will be used to guide the biomarker identification. It is obvious that a better cancer boundary delimitation on the MALDI imaging slice would be beneficial. In this paper, we proposed methods to predict the true cancer boundary, using the MALDI MS data, from the most confident cancer region given by pathologists on an adjacent slice.

Current measurement of the single longest dimension of a polyp is subjective and has variations among radiologists. Our purpose was to develop an automated measurement of polyp volume in CT colonography (CTC). We developed a computerized segmentation scheme for measuring polyp volume in CTC, which consisted of extraction of a highly polyp-like seed region based on the Hessian matrix, segmentation of polyps by use of a 3D volume-growing technique, and sub-voxel refinement to reduce a bias of segmentation. Our database consisted of 30 polyp views (15 polyps) in CTC scans from 13 patients. To obtain "gold standard," a radiologist outlined polyps in each slice and calculated volumes by summation of areas. The measurement study was repeated three times at least one week apart for minimizing a memory effect bias. We used the mean volume of the three studies as "gold standard." Our measurement scheme yielded a mean polyp volume of 0.38 cc (range: 0.15-1.24 cc), whereas a mean "gold standard" manual volume was 0.40 cc (range: 0.15-1.08 cc). The mean absolute difference between automated and manual volumes was 0.11 cc with standard deviation of 0.14 cc. The two volumetrics reached excellent agreement (intra-class correlation coefficient was 0.80) with no statistically significant difference (p(F≤f) = 0.42). Thus, our automated scheme efficiently provides accurate polyp volumes for radiologists.

We have developed a computerized evaluation method of white matter hyperintensity (WMH) regions for the diagnosis of vascular dementia (VaD) based on magnetic resonance (MR) images, and implemented the proposed method as a graphical interface program. The WMH regions were segmented using either a region growing technique or a level set method, one of which was selected by using a support vector machine. We applied the proposed method to MR images acquired from 10 patients with a diagnosis of VaD. The mean similarity index between WMH regions determined by a manual method and the proposed method was 78.2±11.0%. The proposed method could effectively assist neuroradiologists in evaluating WMH regions.

A machine learning technique is presented for assessing brain tumor progression by exploring six patients' complete MRI records scanned during their visits in the past two years. There are ten MRI series, including diffusion tensor image (DTI), for each visit. After registering all series to the corresponding DTI scan at the first visit, annotated normal and tumor regions were overlaid. Intensity value of each pixel inside the annotated regions were then extracted across all of the ten MRI series to compose a 10 dimensional vector. Each feature vector falls into one of three categories:normal, tumor, and normal but progressed to tumor at a later time. In this preliminary study, we focused on the trend of brain tumor progression during three consecutive visits, i.e., visit A, B, and C. A machine learning algorithm was trained using the data containing information from visit A to visit B, and the trained model was used to predict tumor progression from visit A to visit C. Preliminary results showed that prediction for brain tumor progression is feasible. An average of 80.9% pixel-wise accuracy was achieved for tumor progression prediction at visit C.

We study Parkinson's disease (PD) using an automatic specialized diffusion-based atlas. A total of 47 subjects, among who 22 patients diagnosed clinically with PD and 25 control cases, underwent DTI imaging. The EPIs have lower resolution but provide essential anisotropy information for the fiber tracking process. The two volumes of interest (VOI) represented by the Substantia Nigra and the Putamen are detected on the EPI and FA respectively. We use the VOIs for the geometry-based registration. We fuse the anatomical detail detected on FA image for the putamen volume with the EPI. After 3D fibers growing on the two volumes, we compute the fiber density (FD) and the fiber volume (FV). Furthermore, we compare patients based on the extracted fibers and evaluate them according to Hohen&Yahr (H&Y) scale. This paper introduces the method used for automatic volume detection and evaluates the fiber growing method on these volumes. Our approach is important from the clinical standpoint, providing a new tool for the neurologists to evaluate and predict PD evolution. From the technical point of view, the fusion approach deals with the tensor based information (EPI) and the extraction of the anatomical detail (FA and EPI).

Traumatic brain injury (TBI) is a major cause of death and disability. Computed Tomography (CT) scan is widely used in the diagnosis of TBI. Nowadays, large amount of TBI CT data is stacked in the hospital radiology department. Such data and the associated patient information contain valuable information for clinical diagnosis and outcome prediction. However, current hospital database system does not provide an efficient and intuitive tool for doctors to search out cases relevant to the current study case. In this paper, we present the TBIdoc system: a content-based image retrieval (CBIR) system which works on the TBI CT images. In this web-based system, user can query by uploading CT image slices from one study, retrieval result is a list of TBI cases ranked according to their 3D visual similarity to the query case. Specifically, cases of TBI CT images often present diffuse or focal lesions. In TBIdoc system, these pathological image features are represented as bin-based binary feature vectors. We use the Jaccard-Needham measure as the similarity measurement. Based on these, we propose a 3D similarity measure for computing the similarity score between two series of CT slices. nDCG is used to evaluate the system performance, which shows the system produces satisfactory retrieval results. The system is expected to improve the current hospital data management in TBI and to give better support for the clinical decision-making process. It may also contribute to the computer-aided education in TBI.

In this work, we introduce a new representation technique of 2D contour shapes and a sequence similarity measure to characterize 2D regions of interest in medical images. First, we define a distance function on contour points in order to map the shape of a given contour to a sequence of real numbers. Thus, the computation of shape similarity is reduced to the matching of the obtained sequences. Since both a query and a target sequence may be noisy, i.e., contain some outlier elements, it is desirable to exclude the outliers in order to obtain a robust matching performance. For the computation of shape similarity, we propose the use of an algorithm which performs elastic matching of two sequences. The contribution of our approach is that, unlike previous works that require images to be warped according to a template image for measuring their similarity, it obviates this need, therefore it can estimate image similarity for any type of medical image in a fast and efficient manner. To demonstrate our method's applicability, we analyzed a brain image dataset consisting of corpus callosum shapes, and we investigated the structural differences between children with chromosome 22q11.2 deletion syndrome and controls. Our findings indicate that our method is quite effective and it can be easily applied on medical diagnosis in all cases of which shape difference is an important clue.

We propose a novel framework for population analysis of DW-MRI data using the Tubular Surface Model. We focus on the Cingulum Bundle (CB) - a major tract for the Limbic System and the main connection of the Cingulate Gyrus, which has been associated with several aspects of Schizophrenia symptomatology. The Tubular Surface Model represents a tubular surface as a center-line with an associated radius function. It provides a natural way to sample statistics along the length of the fiber bundle and reduces the registration of fiber bundle surfaces to that of 4D curves. We apply our framework to a population of 20 subjects (10 normal, 10 schizophrenic) and obtain excellent results with neural network based classification (90% sensitivity, 95% specificity) as well as unsupervised clustering (k-means). Further, we apply statistical analysis to the feature data and characterize the discrimination ability of local regions of the CB, as a step towards localizing CB regions most relevant to Schizophrenia.

The efficiency of breast density assessment using interactive intensity thresholding applied to intensity uniformity corrected T1-weighted MR images is investigated for 20 healthy women who attended the UK multi-centre study of MRI screening for breast cancer. Mammographic density is estimated on the medial-lateral oblique X-ray mammograms using CUMULUS. MR density assessment is performed using both high and low-resolution T1-weighted images. The left and the right breast regions anterior to the pectoral muscle were segmented on these images using active contouring. For each region, intensity uniformities were corrected using proton density images and a user selected uniformity factor. An interactively selected threshold is applied to the corrected images to detect fibrogulandular tissue. The breast density is calculated as the ratio of the classified fibroglandular tissue to the segmented breast volume. There is no systematic difference, good consistency and a high correlation between the left and the right breast densities estimated from X-ray mammograms and the high and low-resolution MR images. The correlation is the highest and the consistency is the best for the low-resolution MR measurements (r=0.976, MeanAbsoluteDifference = 2.12%). Mean breast densities calculated over the left and the right breasts on high and low-resolution MR images are highly correlated with mammographic density (r=0.923 and 0.903, respectively) but are approximately 50% lower. Interactive intensity thresholding of T1-weighted MR images provides an easy, reproducible and reliable way to assess breast density. High and low-resolution measurements are both highly correlated with the mammographic density but the latter requires less processing and acquisition time.

The purpose was to investigate the repeatability and bias of the output of two classifiers commonly used in computeraided diagnosis for the task of distinguishing benign from malignant lesions. Classifier training and testing were performed within a bootstrap approach using a dataset of 125 sonographic breast lesions (54 malignant, 71 benign). The classifiers investigated were linear discriminant analysis (LDA) and a Bayesian Neural Net (BNN) with 5 hidden units. Both used the same 4 input lesion features. The bootstrap .632plus area under the ROC curve (AUC) was used as a summary performance metric. On an individual case basis, the variability of the classifier output was used in a detailed performance evaluation of repeatability and bias. The LDA obtained an AUC value of 0.87 with 95% confidence interval [0.81; 0.92]. For the BNN, those values were 0.86 and [.76; .93], respectively. The classifier outputs for individual cases displayed better repeatability (less variability) for the LDA than for the BNN and for the LDA the maximum repeatability (lowest variability) lied in the middle of the range of possible outputs, while the BNN was least repeatable (highest variability) in this region. There was a small but significant systematic bias in the LDA output, however, while for the BNN the bias appeared to be weak. In summary, while ROC analysis suggested similar classifier performance, there were substantial differences in classifier behavior on a by-case basis. Knowledge of this behavior is crucial for successful translation and implementation of computer-aided diagnosis in clinical decision making.

Computerized texture analysis of mammographic images has emerged as a means to characterize breast parenchyma and estimate breast percentage density, and thus, to ultimately assess the risk of developing breast cancer. However, during the digitization process, mammographic images may be modified and optimized for viewing purposes, or mammograms may be digitized with different scanners. It is important to demonstrate how computerized texture analysis will be affected by differences in the digital image acquisition. In this study, mammograms from 172 subjects, 30 women with the BRCA1/2 gene-mutation and 142 low-risk women, were retrospectively collected and digitized. Contrast enhancement based on a look-up table that simulates the histogram of a mixed-density breast was applied on very dense and very fatty breasts. Computerized texture analysis was performed on these transformed images, and the effect of variable gain on computerized texture analysis on mammograms was investigated. Area under the receiver operating characteristic curve (AUC) was used as a figure of merit to assess the individual texture feature performance in the task of distinguishing between the high-risk and the low-risk women for developing breast cancer. For those features based on coarseness measures and fractal measures, the histogram transformation (contrast enhancement) showed little effect on the classification performance of these features. However, as expected, for those features based on gray-scale histogram analysis, such as balance and skewnesss, and contrast measures, large variations were observed in terms of AUC values for those features. Understanding this effect will allow us to better assess breast cancer risk using computerized texture analysis.

In breast MRI, intensity inhomogeneity due to coil profile hampers development of robust segmentation and automated processing methods. The purpose of this paper is to evaluate the performance in breast MRI of a number of existing non-uniformity correction methods, mostly developed for brain imaging, and a novel correction method first presented here. Ten breast MRI exams, which were manually segmented into background and five tissue classes, were used for performance assessment. Results show that the relatively simple and fast bias field correction method presented in this paper outperforms the other methods in a number of aspects.

An interactive computer-aided detection or diagnosis (ICAD) scheme allows observers to query suspicious abnormalities (lesions) depicted on medical images. Once a suspicious region is queried, ICAD segments the abnormal region, computes a set of image features, searches for and identifies the reference regions depicted on the verified lesions that are similar to the queried one. Based on the distribution of the selected similar regions, ICAD generates a detection (or classification) score of the queried region depicting true-positive disease. In this study, we assessed the performance and reliability of an ICAD scheme when using a database including total 1500 positive images depicted verified breast masses and 1500 negative images depicted ICAD-cued false-positive regions as well as the leave-one-out testing method. We conducted two experiments. In the first experiment, we tested the relationship between ICAD performance and the size of reference database by systematically increasing the size of reference database from 200 to 3000 images. In the second experiment, we tested the relationship between ICAD performance and the similarity level between the queried image and the retrieved similar references by applying a set of thresholds to systematically remove the queried images whose similarity level to their most "similar" reference images are lower than threshold. The performance was compared based on the areas under ROC curves (AUC). The results showed that (1) as the increase of reference database, AUC value monotonically increased from 0.636±0.041 to 0.854±0.004 and (2) as the increase of similarity threshold values, AUC value also monotonically increased from 0.854±0.004 to 0.932±0.016. The increase of AUC values and the decrease of their standard deviations indicate the improvement of both CAD performance and reliability. The study suggested that (1) assembling the large and diverse reference databases and (2) assessing and reporting the reliability of ICAD-generated results based on the similarity measurement are important in development and application of the ICAD schemes.

This paper presents an automated scheme for breast density estimation on mammogram using statistical and boundary information. Breast density is regarded as a meaningful indicator for breast cancer risk, but measurement of breast density still relies on the qualitative judgment of radiologists. Therefore, we attempted to develop an automated system achieving objective and quantitative measurement. For preprocessing, we first segmented the breast region, performed contrast stretching, and applied median filtering. Then, two features were extracted: statistical information including standard deviation of fat and dense regions in breast area and boundary information which is the edge magnitude of a set of pixels with the same intensity. These features were calculated for each intensity level. By combining these features, the optimal threshold was determined which best divided the fat and dense regions. For evaluation purpose, 80 cases of Full-Field Digital Mammography (FFDM) taken in our institution were utilized. Two observers conducted the performance evaluation. The correlation coefficients of the threshold and percentage between human observer and automated estimation were 0.9580 and 0.9869 on average, respectively. These results suggest that the combination of statistic and boundary information is a promising method for automated breast density estimation.

Mammography reading by radiologists and breast tissue image interpretation by pathologists often leads to high False Positive (FP) Rates. Similarly, current Computer Aided Diagnosis (CADx) methods tend to concentrate more on sensitivity, thus increasing the FP rates. A novel method is introduced here which employs similarity based method to decrease the FP rate in the diagnosis of microcalcifications. This method employs the Principal Component Analysis (PCA) and the similarity metrics in order to achieve the proposed goal. The training and testing set is divided into generalized (Normal and Abnormal) and more specific (Abnormal, Normal, Benign) classes. The performance of this method as a standalone classification system is evaluated in both the cases (general and specific). In another approach the probability of each case belonging to a particular class is calculated. If the probabilities are too close to classify, the augmented CADx system can be instructed to have a detailed analysis of such cases. In case of normal cases with high probability, no further processing is necessary, thus reducing the computation time. Hence, this novel method can be employed in cascade with CADx to reduce the FP rate and also avoid unnecessary computational time. Using this methodology, a false positive rate of 8% and 11% is achieved for mammography and cellular images respectively.

Computer-assisted diagnosis (CADx) for the characterization of mammographic masses as benign or malignant has a very high potential to help radiologists during the critical process of diagnostic decision making. By default, the characterization of mammographic masses is performed by extracting features from a region of interest (ROI) depicting the mass. To investigate the influence of the region on the classification performance, textural, morphological, frequency- as well as moment-based features are calculated in subregions of the ROI, which has been delineated manually by an expert. The investigated subregions are (a) the semi-automatically segmented area which includes only the core of the mass, (b) the outer border region of the mass, and (c) the combination of the outer and the inner border region, referred to as mass margin. To extract the border region and the margin of a mass an extended version of the rubber band straightening transform (RBST) was developed. Furthermore, the effectiveness of the features extracted from the RBST transformed border region and mass margin is compared to the effectiveness of the same features extracted from the untransformed regions. After the feature extraction process a preferably optimal feature subset is selected for each feature extractor. Classification is done using a k-NN classifier. The classification performance was evaluated using the area A_z under the receiver operating characteristic curve. A publicly available mammography database was used as data set. Results showed that the manually drawn ROI lead to superior classification performances for the morphological feature extractors and that the transformed outer border region and the mass margin are not suitable for moment-based features but yield to promising results for textural and frequency-based features. Beyond that the mass margin, which combines the inner and the outer border region, leads to better classification performances compared to the outer border region for its own.

Computer aided diagnosis (CADx) systems can support the radiologist in the complex task of discriminating benign and malignant mammographic lesions. Automatic segmentation of mammographic lesions in regions of interest (ROIs) is a core module of many CADx systems. Previously, we have proposed a novel method for segmentation of mammographic masses. The approach was based on the observation that the optical density of a mass is usually high near its core and decreases towards its boundary. In the work at hand, we improve this approach by integration of a pre-processing module for the correction of inhomogeneous background tissue and by improved selection of the optimal mass contour from a list of candidates based on a cost function. We evaluate the performance of the proposed approach using ten-fold cross-validation on a database of mass lesions and ground-truth segmentations. Furthermore, we compare the improved segmentation approach with the previously proposed approach and with implementations of two state of the art approaches. The results of our study indicate that the proposed approach outperforms both the original method and the two state of the art methods.

A new methodology for computer aided diagnosis in digital mammography using unsupervised classification and classdependent feature selection is presented. This technique considers unlabeled data and provides unsupervised classes that give a better insight into classes and their interrelationships, thus improving the overall effectiveness of the diagnosis. This technique is also extended to utilize biclustering methods, which allow for definition of unsupervised clusters of both pathologies and features. This has potential to provide more flexibility, and hence better diagnostic accuracy, than the commonly used feature selection strategies. The developed methods are applied to diagnose digital mammographic images from the Mammographic Image Analysis Society (MIAS) database and the results confirm the potential for improving the current diagnostic rates.

In this paper a new approach to mass classification based on multi-agent (MA) method is proposed for CAD in mammography. Multi-agent method is used here as a method that fuses the classification information from multiple classifiers in order to obtain a better decision result. Each agent receives the measurement value of individual classifier as initial value in classifying a sample and sends a message to a decision center. The decision center responds to this message with analysis of the correlation among these classifiers and their own decisions information. If the analysis result is conformable to a given standard, the center will provide a final result. Otherwise the message of agent had to be modified iteratively. 128 ROIs, including 64 benign masses and 64 malignant masses, from the DDSM, were used in the mass classification experiment. In comparison with the majority voting based fusion method, we evaluated the performance of proposed multi-agent fusion approach in distinguishing malignant and benign masses. The results demonstrated that the multi-agent method outperforms the majority voting method. Multi-agent fusion method yielded an accuracy of 95.47%, while the majority voting method had an accuracy of 92.23%. In addition, a preliminary study of MA method for mass classification under the bi-view model is reported. All of these experiments showed that the multi-agent method can play a significant role in multiple classifier fusion to improve mass classification in mammography.

In Cone Beam Breast CT (CBBCT), breast calcifications have higher intensities than the surrounding tissues. Without the superposition of breast structures, the three-dimensional distribution of the calcifications can be revealed. In this research, based on the fact that calcifications have higher contrast, a local thresholding and a histogram thresholding were used to select candidate calcification areas. Six features were extracted from each candidate calcification: average foreground CT number value, foreground CT number standard deviation, average background CT number value, background CT number standard deviation, foreground-background contrast, and average edge gradient. To reduce the false positive candidate calcifications, a feed-forward back propagation artificial neural network was designed. The artificial neural network was trained with the radiologists confirmed calcifications and used as classifier in the calcification auto-detection task. In the preliminary experiments, 90% of the calcifications in the testing data sets were detected correctly with an average of 10 false positives per data set.

Recently, tomosynthesis (DBT) and CT (BCT) have been developed for breast imaging. Since each modality produces a fundamentally different representation of the breast volume, our goal was to investigate whether a 3D segmentation algorithm for breast masses could be applied to both DBT and breast BCT images. A secondary goal of this study was to investigate a simplified method for comparing manual outlines to a computer segmentation. The seeded mass lesion segmentation algorithm is based on maximizing the radial gradient index (RGI) along a constrained region contour. In DBT, the constraint function was a prolate spherical Gaussian, with a larger FWHM along the depth direction where the resolution is low, while it was a spherical Gaussian for BCT. For DBT, manual lesion outlines were obtained in the in-focus plane of the lesion, which was used to compute the overlap ratio with the computer segmentation. For BCT, lesions were manually outlined in three orthogonal planes, and the average overlap ratio from the three planes was computed. In DBT, 81% of all lesions were segmented at an overlap ratio of 0.4 or higher, based on manual outlines in one slice through the lesion center. In BCT, 93% of all segmentations achieved an average overlap ratio of 0.4, based on the manual outlines in three orthogonal planes. Our results indicate mass lesions in both BCT and DBT images can be segmented with the proposed 3D segmentation algorithm, by selecting an appropriate set of parameters and after images have undergone specific pre-processing.

Coronary artery disease is the end result of the accumulation of atheromatous plaques within the walls of coronary arteries and is the leading cause of death worldwide. Computed tomography angiography (CTA) has been proved to be very useful for accurate noninvasive diagnosis and quantification of plaques. However, the existing methods to measure the stenosis in the plaques are not accurate enough in mid and distal segments where the vessels become narrower. To alleviate this, we propose a method that consists of three stages namely, automatic extraction of coronary vessels; vessels straightening; lumen extraction and stenosis evaluation. In the first stage, the coronary vessels are segmented using a parametric approach based on circular vessel model at each point on the centerline. It is assumed that centerline information is available in advance. Vessel straightening in the second stage performs multi-planar reformat (MPR) to straighten the curved vessels. MPR view of a vessel helps to visualize and measure the plaques better. On the straightened vessel, lumen and vessel wall are segregated using a nearest neighbor classification. To detect the plaques with severe stenosis in the vessel lumen, we propose a "Diameter Luminal Stenosis" method for analyzing the smaller segments of the vessel. Proposed measurement technique identifies the segments that have plaques and reports the top three severely stenosed segments. Proposed algorithm is applied on 24 coronary vessels belonging to multiple cases acquired from Sensation 64 - slice CT and initial results are promising.

In the last decade intravascular optical coherence tomography has known a tremendous progress. Its high resolution (5- 10μm) allows coronary plaque characterization, vulnerable plaque assessment, and the guidance of intravascular interventions. However, one intravascular OCT sequence contains hundreds of frames, and their interpretation requires a lot of time and energy. Therefore, there is a strong need for automated segmentation algorithms to process this large amount of data. In this article, we present an automated algorithm to extract lumen contours from images obtained with intravascular Optical Coherence Tomography (OCT). Unlike existing methods, our algorithm requires no post- or preprocessing of the image. First, a sliding window is passed on every A-scan to locate the artery tissue, this location being determined from the largest distribution of the grey level values. Once all the tissue is extracted from the image, every segmented A-scan is binarized separately. For a single A-scan, the level of amplitude often varies strongly across the tissue. A global threshold would cause low amplitude parts of the tissue to be considered as belonging to the background. Our solution is to determine local thresholds for every A-scan. That is, instead of having a single global threshold, we allow the threshold itself to smoothly vary across the image. Subsequently, on the binarized image the Prewitt mask is moved from the detected tissue position toward the probe to segment the lumen. The proposed method has been validated qualitatively on images acquired under different conditions without changing any parameter of the algorithm. Experimental results show that the proposed method is accurate and robust to extract lumen borders.

The purpose of our study is the evaluation of an algorithm to determine the physiological relevance of a coronary lesion as seen in a coronary angiogram. The aim is to extract as much as possible information from a standard coronary angiogram to decide if an abnormality, percentage of stenosis, as seen in the angiogram, results in physiological impairment of the blood supply of the region nourished by the coronary artery. Coronary angiography, still the golden standard, is used to determine the cause of angina pectoris based on the demonstration of an important stenose in a coronary artery. Dimensions of a lesion such as length and percentage of narrowing can at present easily be calculated by using an automatic computer algorithm such as Quantitative Coronary Angiography (QCA) techniques resulting in just anatomical information ignoring the physiological relevance of the lesion. In our study we analyze myocardial perfusion images in standard coronary angiograms in rest and in artificial hyperemic phases, using a drug e.g. papaverine intracoronary. Setting a Region of Interest (ROI) in the angiogram without overlying major vessels makes it possible to calculate contrast differences as a function of time, so called time-density curves, in the basal and hyperemic phases. In minimizing motion artifacts, end diastolic images are selected ECG based in basal and hyperemic phase in an identical ROI in the same angiographic projection. The development of new algorithms for calculating differences in blood supply in the region as set are presented together with the results of a small clinical case study using the standard angiographic procedure.

Thoracic Aortic Aneurysm (TAA) is a localized swelling of the thoracic aorta. The progressive growth of an aneurysm may eventually cause a rupture if not diagnosed or treated. This necessitates the need for an accurate measurement which in turn calls for the accurate segmentation of the aneurysm regions. Computer Aided Detection (CAD) is a tool to automatically detect and segment the TAA in the Computer tomography angiography (CTA) images. The fundamental major step of developing such a system is to develop a robust method for the detection of main vessel and measuring its diameters. In this paper we propose a novel adaptive method to simultaneously segment the thoracic aorta and to indentify its center line. For this purpose, an adaptive parametric 3D region growing is proposed in which its seed will be automatically selected through the detection of the celiac artery and the parameters of the method will be re-estimated while the region is growing thorough the aorta. At each phase of region growing the initial center line of aorta will also be identified and modified through the process. Thus the proposed method simultaneously detect aorta and identify its centerline. The method has been applied on CT images from 20 patients with good agreement with the visual assessment by two radiologists.

We present a simple rendering scheme for thoracic CT datasets which yields a color coding based on local differential geometry features rather than Hounsfield densities. The local curvatures are computed on several resolution scales and mapped onto different colors, thereby enhancing nodular and tubular structures. The rendering can be used as a navigation device to quickly access points of possible chest anomalies, in particular lung nodules and lymph nodes. The underlying principle is to use the nodule enhancing overview as a possible alternative to classical CAD approaches by avoiding explicit graphical markers. For performance evaluation we have used the LIDC-IDRI lung nodule data base. Our results indicate that the nodule-enhancing overview correlates well with the projection images produced from the IDRI expert annotations, and that we can use this measure to optimize the combination of differential geometry filters.

A computerized image analysis technology suffers from imperfection, imprecision and vagueness of the input data and its propagation in all individual components of the technology including image enhancement, segmentation and pattern recognition. Furthermore, a Computerized Medical Image Analysis System (CMIAS) such as computer aided detection (CAD) technology deals with another source of uncertainty that is inherent in image-based practice of medicine. While there are several technology-oriented studies reported in developing CAD applications, no attempt has been made to address, model and integrate these types of uncertainty in the design of the system components, even though uncertainty issues directly affect the performance and its accuracy. In this paper, the main uncertainty paradigms associated with CAD technologies are addressed. The influence of the vagueness and imprecision in the classification of the CAD, as a second reader, on the validity of ROC analysis results is defined. In order to tackle the problem of uncertainty in the classification design of the CAD, two fuzzy methods are applied and evaluated for a lung nodule CAD application. Type-1 fuzzy logic system (T1FLS) and an extension of it, interval type-2 fuzzy logic system (IT2FLS) are employed as methods with high potential for managing uncertainty issues. The novelty of the proposed classification methods is to address and handle all sources of uncertainty associated with a CAD system. The results reveal that IT2FLS is superior to T1FLS for tackling all sources of uncertainty and significantly, the problem of inter and intra operator observer variability.

We developed an automated method for the segmentation of lungs with severe diffuse interstitial lung disease (DILD) in multi-detector CT. In this study, we would like to compare the performance levels of this method and a thresholdingbased segmentation method for normal lungs, moderately abnormal lungs, severely abnormal lungs, and all lungs in our database. Our database includes 31 normal cases and 45 abnormal cases with severe DILD. The outlines of lungs were manually delineated by a medical physicist and confirmed by an experienced chest radiologist. These outlines were used as reference standards for the evaluation of the segmentation results. We first employed a thresholding technique for CT value to obtain initial lungs, which contain normal and mildly abnormal lung parenchyma. We then used texture-feature images derived from co-occurrence matrix to further segment lung regions with severe DILD. The segmented lung regions with severe DILD were combined with the initial lungs to generate the final segmentation results. We also identified and removed the airways to improve the accuracy of the segmentation results. We used three metrics, i.e., overlap, volume agreement, and mean absolute distance (MAD) between automatically segmented lung and reference lung to evaluate the performance of our segmentation method and the thresholding-based segmentation method. Our segmentation method achieved a mean overlap of 96.1%, a mean volume agreement of 98.1%, and a mean MAD of 0.96 mm for the 45 abnormal cases. On the other hand the thresholding-based segmentation method achieved a mean overlap of 94.2%, a mean volume agreement of 95.8%, and a mean MAD of 1.51 mm for the 45 abnormal cases. Our new method obtained higher performance level than the thresholding-based segmentation method.

The development of computer-aided diagnosis (CAD) methods for the processing of CT lung scans continues to become increasingly popular due to the potential of these algorithms to reduce image reading time, errors caused by user fatigue, and user subjectivity when screening for the presence of malignant lesions. This study seeks to address the critical need for a realistic simulated lung nodule CT image dataset based on real tumor morphologies that can be used for the quantitative evaluation and comparison of these CAD algorithms. The manual contouring of 17 different lung metastases was performed and reconstruction of the full 3-D surface of each tumor was achieved through the utilization of an analytical equation comprised of a spherical harmonics series. 2-D nodule slice representations were then computed based on these analytical equations to produce realistic simulated nodules that can be inserted into CT datasets with well-circumscribed, vascularized, or juxtapleural borders and also be scaled to represent nodule growth. The 3-D shape and intensity profile of each simulated nodule created from the spherical harmonics reconstruction was compared to the real patient CT lung metastasis from which its contour points were derived through the calculation of a 3-D correlation coefficient, producing an average value of 0.8897 (±0.0609). This database of realistic simulated nodules can fulfill the need for a reproducible and reliable gold standard for CAD algorithms with regards to nodule detection and sizing, especially given its virtually unlimited capacity for expansion to other nodule shape variants, organ systems, and imaging modalities.

Computer-aided diagnostic characterization (CADc) aims to support medical imaging decision making by objectively rating the radiologists' subjective, perceptual opinions of visual diagnostic characteristics of suspicious lesions. This research uses the publicly available Lung Image Database Consortium (LIDC) collection of radiologists' outlines of nodules and ratings of boundary and shape characteristics: spiculation, margin, lobulation, and sphericity. The approach attempts to reduce the observed disagreement between radiologists on the extent of nodules by combining their spatial opinion using probability maps to create regions of interest (ROIs). From these ROIs, images features are extracted and combined using machine learning models to predict a combined opinion, the median rating and a thresholded, binary version of their diagnostic characteristics. The results show slight to fair agreement-linear-weighted Kappa-between the CADc models and median radiologist opinion for the full scale five-level rating and fair to moderate agreement using a binary version of the median radiologist opinion.

In this preliminary study, a new computer-aided detection (CAD) scheme for pulmonary embolism (PE) detection was developed and tested. The scheme applies multiple steps including lung segmentation, candidate extraction using intensity mask and tobogganing method, feature extraction, false positive reduction using a multifeature based artificial neural network (ANN) and a k-nearest neighbor (KNN) classifier to detect and classify suspicious PE lesions. In particular, a new method to define the surrounding background regions of interest (ROI) depicting PE candidates was proposed and tested in an attempt to reduce the detection of false positive regions. In this study, the authors also investigated following methods to improve CAD performance, which include a grouping and scoring method, feature selection using genetic algorithm, and limitation on allowed suspicious lesions to be cued in one examination. To test the scheme performance, a set of 20 chest CT examinations were selected. Among them, 18 are positive cases depicted 44 verified PE lesions and the remaining 2 were negative cases. The dataset was also divided into a training subset (9 examinations) and a testing subset (11 examinations), respectively. The experimental results showed when applying to the testing dataset CAD scheme using tobogganing method alone achieved 2D region-based sensitivity of 72.1% (220/305) and 3D lesion-based sensitivity of 83.3% (20/24) with total 19,653 2D false-positive (FP) PE regions (1,786.6 per case or approximately 6.3 per CT slice). Applying the proposed new method to improve lung region segmentation and better define the surrounding background ROI, the scheme reduced the region-based sensitivity by 6.5% to 65.6% or lesion-based sensitivity by 4.1% to 79.2% while reducing the FP rate by 65.6% to 6,752 regions (or 613.8 per case). After applying the methods of grouping, the maximum scoring, a genetic algorithm (GA) to delete "redundant" features, and limiting the maximum number of cued-lesions in one examination, CAD scheme further reduced FP rate to 50 per case. Based on the FROC curve, an operating threshold was set up in which the CAD scheme could ultimately achieve 63.2% detection sensitivity with 18.4 FP regions per case when applying to the testing dataset. This study investigated the feasibility of several methods applying to the CAD scheme in detecting PE lesions and demonstrated that CAD performance could depend on many factors including better defining candidate ROI and its background, optimizing the 2D region grouping and scoring methods, selecting the optimal feature set, and limiting the number of allowed cueing lesions per examination.

Computer-Aided Detection (CAD) systems are becoming widespread supporting tools to radiologists' diagnosis, especially in screening contexts. However, a large amount of false positive (FP) alarms would inevitably lead both to an undesired possible increase in time for diagnosis, and to a reduction in radiologists' confidence in CAD as a useful tool. Most CAD systems implement as final step of the analysis a classifier which assigns a score to each entry of a list of findings; by thresholding this score it is possible to define the system performance on an annotated validation dataset in terms of a FROC curve (sensitivity vs. FP per scan). To use a CAD as a supportive tool for most clinical activities, an operative point has to be chosen on the system FROC curve, according to the obvious criterion of keeping the sensitivity as high as possible, while maintaining the number of FP alarms still acceptable. The strategy proposed in this study is to choose an operative point with high sensitivity on the CAD FROC curve, then to implement in cascade a further classification step, constituted by a smarter classifier. The key issue of this approach is that the smarter classifier is actually a meta-classifier of more then one decision system, each specialized in rejecting a particular type of FP findings generated by the CAD. The application of this approach to a dataset of 16 lung CT scans previously processed by the ^VBNACAD system is presented. The lung CT ^VBNACAD performance of 87.1% sensitivity to juxtapleural nodules with 18.5 FP per scan is improved up to 10.1 FP per scan while maintaining the same value of sensitivity. This work has been carried out in the framework of the MAGIC-V collaboration.