Computer-aided diagnostic (CAD) schemes have been developed to assist radiologists in the early detection of lung cancer in radiographs and computed tomography (CT) images. In order to improve the sensitivity for nodule detection, many researchers have employed a filter as a preprocessing step for enhancement of nodules. However, these filters enhance not only nodules, but also other anatomic structures such as ribs, blood vessels, and airway walls. Therefore, nodules are often detected together with a large number of false positives caused by these normal anatomic structures. In this study, we developed a selective enhancement filter for nodule which can simultaneously enhance nodules and suppress other normal anatomic structures such as blood vessels and airway walls. Therefore, as preprocessing steps, this filter would be useful for improving the sensitivity of nodule detection and for reducing the number of false positives. We applied the enhancement filter to two-dimensional (2-D) and three-dimensional (3-D) CT images and also to 3-D magnetic resonance angiography (MRA) images to show its effectiveness in the enhancement of lung nodules and intracranial aneurysms.

Lung nodules that exhibit growth over time are considered highly suspicious for malignancy. We present a completely automated system for detection of growing lung nodules, using initial and follow-up multi-slice CT studies. The system begins with automatic detection of lung nodules in the later CT study, generating a preliminary list of candidate nodules. Next an automatic system for registering locations in two studies matches each candidate in the later study to its corresponding position in the earlier study. Then a method for automatic segmentation of lung nodules is applied to each candidate and its matching location, and the computed volumes are compared. The output of the system is a list of nodule candidates that are new or have exhibited volumetric growth since the previous scan. In a preliminary test of 10 patients examined by two radiologists, the automatic system identified 18 candidates as growing nodules. 7 (39%) of these corresponded to validated nodules or other focal abnormalities that exhibited growth. 4 of the 7 true detections had not been identified by either of the radiologists during their initial examinations of the studies. This technique represents a powerful method of surveillance that may reduce the probability of missing subtle or early malignant disease.

Helical computed tomography (CT) is the premier modality for chest imaging and is increasingly used to detect lung nodules and quantify their size. Several automated methods have been developed to maximize the precision of volume measurement of a nodule once it is identified. Binary variable threshold (VT) and the partial volume (PV) are two commonly used techniques for estimating nodule volume. We present and validate on a phantom an algorithm that combines the benefits of VT and PV techniques. The process begins by defining an initial binary approximation, Wi, The hybrid algorithm then constructs a hollow shell dW around the edges. The shell dW is designed to maximize the likelihood of containing the lung-nodule interface, while minimizing the containment of both the core of nodule and the surrounding lung. Finally, the PV method is applied to dW and the resulting volume is added to the interior volume to obtain the volume of the entire nodule. We have implemented the hybrid method and validated it on a realistic chest phantom that contained 40 simulated nodules. The absolute error of VT and PV was 2.3 mm³, with a standard deviation of 2.0 mm³. The hybrid method performed better, with an absolute error of 2.2 mm³ and a standard deviation of 1.9 mm³. The algorithm appears to be a promising tool in evaluating the growth of lung nodule.

The purpose of this study is to develop a computer-aided diagnosis (CAD) system to detect small-sized (from 2mm to 10mm) non-pleural pulmonary nodules in high resolution helical CT scans. A new 3D automated scheme using cellular neural networks is presented. Different from most previous methods, this scheme employed the local shape property to perform voxel classification. The shape index feature successfully captured the local shape difference between nodules and non-nodules, especially vessels. A 3D discrete-time cellular neural network (DTCNN) was constructed to give a reliable voxel classification by collecting information in a neighborhood. To tailor it for lung nodule detection, this DTCNN was trained using genetic algorithms (GAs) to derive the shape index variation pattern of nodules. 19 clinical thoracic CT cases involving a total of 4838 sectional images were used in this work, with 2 scans forming the training set, and the remaining 17 cases being the testing set. The evaluation was composed of two stages. During the first stage, a pulmonologist and our CAD system independently detected nodules in the testing set. Then, the suspected nodule areas located by the computer were reviewed by the pulmonologist to confirm nodules missed by the human in the first review. There were 32 true nodules detected by the computer but missed by the pulmonologist in the first review, in which 30 non-juxtapleural nodules were found. Considering the nodules detected by the pulmonologist during the first and second reviews as the truth, 52 of 62 non-pleural nodules were detected by the CAD system (sensitivity being 83.9%), with the number of false positives being 3.47 per case.

Emphysema is a common chronic respiratory disorder characterized by the destruction of lung tissue. Labelling lung images containing Emphysema is a tedious and time consuming process and detection using fewer labelled examples would be very useful. We have recently developed an automated texture-based system capable of achieving varying levels of Emphysema detection in High Resolution Computed Tomography (HRCT) images using co-training [1]. Co-training is a semi-supervised technique used to improve classifiers trained with very few labelled examples using a large pool of unseen examples. In this paper, we show how we can use examples labelled by experts within the same system but in an incremental manner. We show that through the use of two views in our system, one can detect the most informative examples and through the use of labelled examples, one can provide class labels to those examples. When the two views disagree about the class label of an example, we feed the example together with the correct class label provided by the expert’s marking to the system in order to improve its performance. Results show that when images labelled by experts are incorporated into the system at early iterations, the performance of the system compared to the earlier system improves. The results were also compared against "density mask", a standard approach used for Emphysema detection in medical image analysis. In addition, radiologists have verified the results and concluded that the classifiers built at different iterations can be used for different levels of Emphysema diagnosis.

We have previously evaluated the effects of computer-aided diagnosis (CAD) on radiologists' characterization of malignant and benign breast masses in single-view serial mammograms. In this study, we conducted observer performance experiments with ROC methodology in which the radiologists read the serial mammograms in two-views (CC and MLO) without and with CAD. 47 temporal pairs of two-view serial mammograms (27 malignant and 20 benign) containing masses were chosen from 39 patient files and digitized. The corresponding masses on each temporal pair were analyzed by the CAD program. For this data set, the computer classifier achieved a test A_z value of 0.90. Five MQSA radiologists assessed the two-view temporal pairs and provided estimates of the likelihood of malignancy without and then with CAD. For the five radiologists, the average A_z was 0.81 (range: 0.72-0.88) without CAD and improved to 0.88 (range: 0.86-0.90) with CAD. The improvement was statistically significant (p=0.038). In comparison, the test A_z value of the computer classifier for single view analysis was 0.87. The average A_z of the radiologists for reading the single view temporal pairs without CAD was 0.78 (range: 0.73-0.83) and was improved significantly (p=0.002) to 0.84 (range: 0.77-0.88) with CAD. CAD using interval change analysis can significantly improve radiologists' accuracy in classification of masses. Classification based on information from two-views is more accurate than that based on single view for both the radiologists and the computer classifier. CAD can further improve radiologists' performance even in two-view reading.

Previously, we presented an information-theoretic image retrieval scheme for computer-assisted detection (CAD) of masses in screening mammograms. The purpose of this study is to evaluate the performance of the CAD scheme in the detection of architectural distortion (AD) in mammograms of breasts with different parenchymal densities. The study was based on the Digital Database for Screening Mammography (DDSM). Initially, a databank of 1,337 mammographic regions with confirmed pathology was created. They were all 512x512 pixel regions of interest (ROIs) depicting a mixture of mass findings such as masses (599), asymmetric densities (30), and architectural distortion (52). In addition, there were 656 ROIs depicting normal parenchyma. The CAD scheme was applied as follows. When a suspicious mammographic region was presented to the CAD system for evaluation, the system searched the databank of mammographic regions with known pathology. A decision index was calculated based on the query's similarity with the stored cases. Mutual information was used as the similarity metric. ROC analysis was performed to determine how well the CAD scheme detected the architectural distortion cases. Based on a leave-one out sampling scheme, the overall ROC performance of the CAD system was 0.82±0.03 for the detection of AD in screening mammograms. Specifically, the CAD system achieved perfect performance in fatty breasts (Az=1.0), Az=0.89±0.03 in fibroglandular breasts, Az=0.78±0.05 in heterogeneous breasts, and Az=0.55±0.09 in dense breasts. The CAD scheme performed well in the detection of architectural distortion in screening mammograms. The performance of the CAD scheme deteriorated significantly as breast density increased.

Mammography and ultrasound (US) are two low-cost modalities that are commonly used by radiologists for evaluating breast masses and making biopsy recommendations. The goal of this study was to investigate computerized methods for combining information from these two modalities for mass characterization. Our data set consisted of 3D US images and mammograms of biopsy-proven solid breast masses from 60 patients. Thirty of the masses were malignant and 30 were benign. The US volume was obtained by scanning with an experimental 3D US image acquisition system. After computerized feature extraction from the 3D US images and mammograms, we investigated three methods (A, B and C) for combining the image features or classifier scores from different mammographic views and the US volumes. The classifier scores were analyzed using the receiver operating characteristic (ROC) methodology. The area Az under the ROC curve of the classifier based on US alone was 0.88±0.04 for testing Two classifiers were designed using the mammograms alone, with test Az values of 0.85±0.05 and 0.87±0.05, respectively. The test accuracy of combination methods A, B, and C were 0.89±0.04, 0.92±0.03, and 0.93±0.03, respectively. Our results indicate that combining the image features or classifier scores from the US and mammographic classification methods can improve the accuracy of computerized mass characterization.

This paper describes a computer-assisted algorithm to automatically assess mammographic breast density. The algorithm was applied to 160 cranio-caudal DDSM mammograms (80 Lumisys and 80 Howtek images). The breast region was first segmented from its background using our self-organizing map (SOM) with knowledge-based refinement algorithm (presented previously). A different SOM neural network was subsequently developed to operate within the determined breast region. Multiscale feature vectors from the breast region were used to train the new SOM. The weight vectors of the SOM were then clustered by the K-means method, resulting in a breast region segmented into K different clusters. The prevalence of SOM clusters containing dense tissues was calculated to develop a summary density index. Statistical analysis was applied to optimize the implementation parameters of the summary index. The average summary index was higher in dense breasts than in non-dense breasts. The trend was consistent for both digitizers, though the results were statistically significant for only the Lumisys set. Unsupervised clustering and segmentation of mammograms is a promising approach for automated breast density assessment.

An important question in mammographic image analysis is the importance of the projected view of the breast. Can temporal changes in density be detected equally well using either one of the commonly available views Medio-Lateral (ML) and Cranio-Caudal (CC) or a combination of the two? Two sets of mammograms of 50 patients in a double-blind, placebo controlled hormone replacement therapy (HRT) experiment were used. One set of ML and CC view from 1999 and one from 2001. HRT increases density which means that the degree of separation of the populations (one group receiving HRT and the other placebo) can be used as a measure of how much density change information is carried in a particular view or combination of views. Earlier results have shown a high correlation between CC and ML views leading to the conclusion that only one of them is needed for density assessment purposes. A similar high correlation coefficient was observed in this study (0.85), while the correlation between changes was a bit lower (0.71). Using both views to separate the patients receiving hormones from the ones receiving placebo increased the area under corresponding ROC curves from 0.76 ± 0.04 to 0.79 ± 0.04.

The object of this research is to investigate a method for determining whether two different imaged presentations of a lesion actually represent the same abnormality. Tasks of this kind are expected to arise in numerous applications, including when interpreting multi-view mammograms or multi-modal images for breast cancer diagnosis and when comparing breast images in longitudinal studies for evaluation of disease prognosis or treatment outcome. Currently, we consider the above-stated discrimination task in two-view breast sonography. We are studying image-based feature(s) and are developing general correlation formulation that uses the identified feature(s). By using a database of 262 actual breast lesions we have obtained promising initial results that yield an A_z value of 0.82 in the task of distinguishing between corresponding lesion pairs and non-corresponding lesion pairs.

Detection of stellate patterns is a very important step in computer-aided detection schemes designed for mammography. We introduce a new way of finding these regions based on the use of quadrature filters. The method allows extraction of a certainty measure for each orientation estimate. This makes the method of finding the areas the spicules seem to emanate from more robust than simply basing it on the orientation estimates themselves. The local phase extracted from the filter outputs allows us to discriminate between orientation estimates from edges and dark lines from those generated by bright line structures, i.e. spicules. This makes the method more specific. We also show how the method can be modified for finding non-spiculated masses in digitized mammograms.

Two new similarity measures for rigid image registration, based on the normalization of Jensen's difference applied to Renyi and Tsallis-Havrda-Charvat entropies, are introduced. One measure is normalized by the first term of Jensen's difference, which in our proposal coincides with the marginal entropy, and the other by the joint entropy. These measures can be seen as an extension of two measures successfully applied in medical image registration: the mutual information and the normalized mutual information. Experiments with various registration modalities show that the new similarity measures are more robust than the normalized mutual information for some modalities and a determined range of the entropy parameter. Also, a certain improvement on accuracy can be obtained for a different range of this parameter.

We report an algorithm for automatic elastic registration of whole body computed tomography (CT) and positron emission tomography (PET) images, which would allow accurate localization of viable tumors visible clearly in the perfusion PET image with respect to the anatomy that is better delineated in the CT image. The algorithm generates an elastic transformation field (TA) from a unique quaternion-based interpolation of multiple rigid-body registrations (based on maximization of normalized mutual information), each obtained for hierarchically subdivided image blocks. Fifteen pairs of clinically acquired CT and PET scans were registered using our algorithm. All image pairs registered visually correct. For validation, a set of anatomic landmarks was identified independently by two clinical experts in both images of each CT-PET pair. Based on each expert's marked points, two thin-plate spline-based deformation fields (TE₁, TE₂) were determined for each image pair. Interobserver variability calculated as the mean difference between transformed locations obtained using each of 'TE₁', 'TE₂' and 'TA' (5.8 mm) was comparable to the interobserver variability calculated as the mean difference in transformed locations obtained using only 'TE₁' and 'TE₂' (4.5 mm), suggesting that the accuracy of the registration algorithm is comparable to that of the experts.

Due to their complementary information content, both X-ray computed tomography (CT) and magnetic resonance (MR) imaging are employed in certain clinical cases to improve the understanding of the pathology. To spatially relate the two datasets, image-registration and image fusion is employed. However, registration errors, either global or local, are common and are non-uniform within the image volume. In this paper, we propose a new algorithm, which assesses the quality of the registration locally within the CT-MR volume and provides visual, color-coded feedback to the user about the location and extent of good/bad correspondence between the two images. The proposed registration assessment algorithm is based on a correspondence analysis of bone structures in the CT and MR images. For that purpose, a custom segmentation algorithm for bone in MR images has been developed, which is based on a stochastic threshold computation method. This segmentation method for MR images and the CT-MR registration assessment algorithm were validated on simulated MR datasets and real CT-MR image pairs of the head. Some partial-volume effects occur at the borders of the bone structures and at the bone interfaces with air, which cannot be separated from bone in the MR image. The presented assessment method of CT-MR image registration offers the user a new tool to evaluate the overall and local quality of the registration. With this information, the user does not have to blindly trust the fused CT-MR datasets but can easily identify areas of inaccurate correspondence. The application of the algorithm is so far limited to T1-weighted MR and CT images of the head area.

An image registration quality evaluator (RQE) is proposed to automatically quantify the accuracy of registrations. The RQE, based on an adaptive pattern classifier, is generated from a pair of reference and target images. It is unique to each patient, anatomical site and imaging modality. RQE is applied to patient positioning in cranial radiotherapy using portal/portal and portal/DRR registrations. We adopted 1mm translation and 1° rotation as the maximal acceptable registration errors, reflecting typical clinical setup tolerances. RQE is used to determine the acceptability of a registration. The performance of RQE was evaluated using phantom images containing radio-opaque fiducial markers. Using receiver operating characteristic (ROC) analysis, we estimated the sensitivity and the specificity of the RQE are 0.95 (with 0.89-0.98 confidence interval (CI) at 95% significance level) and 0.95 (with 0.88-0.98 CI at 95% significance level) respectively for intramodal RQE. For intermodal RQE, the sensitivity and the specificity are 0.92 (with 0.81-0.98 CI at 95% significance level) and 0.98 (with 0.89-0.99 CI at 95% significance level) respectively. Clinical use of RQE could significantly reduce the involvement of the oncologist for routine pre-treatment patient positioning verification, while increasing setup accuracy.

In the last decade information-theoretic similarity measures, especially mutual information and its derivatives, have proven to be accurate measures for rigid and non-rigid, mono- and multi-modal image registration. However, these measures are sometimes not robust enough, especially in cases of poor image quality. This is most likely due to the lack of spatial information included in the measure as usually only intensities are employed to measure similarity between images. Spatial information in the form of intensity gradients or second derivatives may be included in information-theoretic similarity measures. This paper presents a novel method for efficiently combining multiple features into the estimation of mutual information. The proposed measure, under certain assumptions on feature probability distribution, strictly follows information theory in contrast to a number of heuristic methods that were proposed to include spatial information in mutual information. The novel approach solves the problem of efficient estimation of multi-feature mutual information from sparse high-dimensional histograms. The proposed measure was tested on widely used Vanderbilt image database. Results indicate that multi-feature mutual information outperforms the single-feature mutual information measure. The contribution of additional image features to registration is especially significant in cases when standard mutual information measure fails. Moreover, it is expected that non-rigid registration may also benefit from the proposed multi-feature mutual information measure.

Automatic polyp detection is a challenging task as polyps come in different sizes and shape. The detection generally consists of colon segmentation, identification of suspected polyps and classification. Classification involves discriminating polyps from among many suspected regions based on a number of features extracted from the detected regions. This paper presents the work on the first two stages of the detection. For the colon segmentation, the fuzzy connectivity region growing technique is used while for the identification of suspected polyps concave region searching is applied. A rule-based filtering based on 3D volumetric features is used to reduce a large number of non-polyp structures (false positives). The method is fast, robust and validated with a number of high-resolution colon datasets.

In coronary x-ray angiographies, the vessels supplying the heart are imaged in a number of states uniquely determined by a combination of the respiratory intake and the heart contraction of the patient. The angiographic frames of one sequence represent not all possible combinations of respiration and heart contraction. A couple of applications need a continuous and dense sampling of the state-space given by the two axes 'respiration' and 'contraction', e.g. background removal or motion-compensated catheter navigation. We present a novel method of interpolating above the twodimensional phase-space based on pairs of angiographic frames with similar contraction, but different respiration status. First a hypothetical model of the respiration motion is formulated, e.g. rigid transformation or rigid translation. Then the parameters that transform a single frame into another one with similar contraction status are calculated for a number of frames. An iterative approach is used to reconstruct the generalized transformation function from the transformation parameters of frame pairs. Using this function, angiographic frames of arbitrary respiration status can be generated. It is shown that the synthesized angiographies closely match real angiographies acquired at the same combination of contraction and respiration status.

Multi-dimensional convex objects can be characterized with respect to shape, structure, and the connectivity of their components using a set of morphological descriptors known as the Minkowski functionals. In a 3D Euclidian space, these correspond to volume, surface area, mean integral curvature, and the Euler-Poincaré characteristic. We introduce the Minkowski functionals to medical image processing for the morphological analysis of trabecular bone tissue. In the context of osteoporosis-a metabolic disorder leading to a weakening of bone due to deterioration of micro-architecture-the structure of bone increasingly gains attention in the quantification of bone quality. The trabecular architecture of healthy cancellous bone consists of a complex 3D system of inter-connected mineralised elements whereas in osteoporosis the micro-structure is dominated by gaps and disconnections. At present, the standard parameter for diagnosis and assessment of fracture risk in osteoporosis is the bone mineral density (BMD) - a bulk measure of mineralisation irrespective of structural texture characteristics. With the development of modern imaging modalities (high resolution MRI, micro-CT) with spatial resolutions allowing to depict individual trabeculae bone micro-architecture has successfully been analysed using linear, 2- dimensional structural measures adopted from standard histo-morphometry. The preliminary results of our study demonstrate that due to the complex - i.e. the non-linear - network of trabecular bone structures non-linear measures in 3D are superior to linear ones in predicting mechanical properties of trabecular bone from structural information extracted from high resolution MR image data.

Cervical cancer is the second most common malignancy in women worldwide. If diagnosed in the premalignant stage, cure is invariably assured. Although the Papanicolaou (Pap) smear has significantly reduced the incidence of cervical cancer where implemented, the test is only moderately sensitive, highly subjective and skilled-labor intensive. Newer optical screening tests (cervicography, direct visual inspection and speculoscopy), including fluorescent and reflective spectroscopy, are fraught with certain weaknesses. Yet, the integration of optical probes for the detection and discrimination of cervical neoplasia with automated image analysis methods may provide an effective screening tool for early detection of cervical cancer, particularly in resource poor nations. Investigative studies are needed to validate the potential for automated classification and recognition algorithms. By applying image analysis techniques for registration, segmentation, pattern recognition, and classification, cervical neoplasia may be reliably discriminated from normal epithelium. The National Cancer Institute (NCI), in cooperation with the National Library of Medicine (NLM), has embarked on a program to begin this and other similar investigative studies.

We introduce the scaling vector method (SVM): a technique to obtain local non-linear structural information from data sets suitable in cases where anisotropy plays an important role. We apply the SVM to High Resolution Magnetic Resonance Images (HRMRI) of human proximal femur specimens IN VITRO which contain a large portion of bone tissue forming a complex network of mineralized trabeculae oriented along the major stress lines. By means of the SVM, we extract a 3D non-linear local anisotropic texture measure which we use to compare with similar isotropic texture measures, bone mineral density (BMD) and standard isotropic 2D (linear) morphometric parameters in the prediction of the biomechanical properties of the trabecular bone. Our results show that structural non-linear anisotropic texture measures which account for the preferential direction of the trabeculae are superior to isotropic and bulk measures in the prediction of the mechanical properties of the human proximal femur in vitro.

The scaling index method (SIM) is a recently proposed non-linear technique to extract texture measures for the quantitative characterisation of the trabecular bone structure in high resolution magnetic resonance imaging (HR-MRI). The three-dimensional tomographic images are interpreted as a point distribution in a state space where each point (voxel) is defined by its x, y, z coordinates and the grey value. The SIM estimates local scaling properties to describe the nonlinear morphological features in this four-dimensional point distribution. Thus, it can be used for differentiating between cluster-, rod-, sheet-like and unstructured (background) image components, which makes it suitable for quantifying the microstructure of human cancellous bone. The SIM was applied to high resolution magnetic resonance images of the distal radius in patients with and without osteoporotic spine fractures in order to quantify the deterioration of bone structure. Using the receiver operator characteristic (ROC) analysis the diagnostic performance of this texture measure in differentiating patients with and without fractures was compared with bone mineral density (BMD). The SIM demonstrated the best area under the curve (AUC) value for discriminating the two groups. The reliability of our new texture measure and the validity of our results were assessed by applying bootstrapping resampling methods. The results of this study show that trabecular structure measures derived from HR-MRI of the radius in a clinical setting using a recently proposed algorithm based on a local 3D scaling index method can significantly improve the diagnostic performance in differentiating postmenopausal women with and without osteoporotic spine fractures.

In neuroimaging research, shape analysis has become a field of great interest due to the ability to locate morphological brain changes between different groups. Currently, many local shape analysis approaches fail to correct for their high number of correlated statistical tests. This can result in an overly optimistic estimate of the local shape analysis. This paper presents a correction scheme for objects described by the parametrized 3D closed surface description SPHARM. The SPHARM parameterization was determined via an area preserving, distortion minimizing optimization. The correction scheme decomposes the object surface into overlapping planar images via a cylindrical equal area projection of the parameterization. The images are individually analyzed with the SnPM/SPM package using a voxel-level non-parametric multiple testing procedure based on permutation tests. The correction scheme employs conservative tests resulting in a pessimistic estimate. We present an application of the correction scheme to the shape similarity analysis of lateral ventricles

We developed a general method for global conformal parameterizations based on the structure of the cohomology group of holomorphic one-forms with or without boundaries. For genus zero surfaces, our algorithm can find a unique mapping between any two genus zero manifolds by minimizing the harmonic energy of the map. In this paper, we apply the algorithm to the cortical surface matching problem. We use a mesh structure to represent the brain surface. Further constraints are added to ensure that the conformal map is unique. Empirical tests on MRI data show that the mappings preserve angular relationships, are stable in MRIs acquired at different times, and are robust to differences in data triangulation, and resolution. Compared with other brain surface conformal mapping algorithms, our algorithm is more stable and has good extensibility.

We address the issue of modeling and quantifying myocardial contraction from 4D MR sequences, and present an unsupervised approach for building and using a statistical 3D motion atlas for the normal heart. This approach relies on a state-of-the-art variational non rigid registration (NRR) technique using generalized information measures, which allows for robust intra-subject motion estimation and inter-subject anatomical alignment. The atlas is built from a collection of jointly acquired tagged and cine MR exams in short- and long-axis views. Subject-specific non parametric motion estimates are first obtained by incremental NRR of tagged images onto the end-diastolic (ED) frame. Individual motion data are then transformed into the coordinate system of a reference subject using subject-to-reference mappings derived by NRR of cine ED images. Finally, principal component analysis of aligned motion data is performed for each cardiac phase, yielding a mean model and a set of eigenfields encoding kinematic ariability. The latter define an organ-dedicated hierarchical motion basis which enables parametric motion measurement from arbitrary tagged MR exams. To this end, the atlas is transformed into subject coordinates by reference-to-subject NRR of ED cine frames. Atlas-based motion estimation is then achieved by parametric NRR of tagged images onto the ED frame, yielding a compact description of myocardial contraction during diastole.

All biological objects exhibit some degree of asymmetry, but for some parts of the human body, excessive asymmetry is a sign of pathology. Hence, the problem is to draw the line between categorization of objects being too asymmetric and objects exhibiting normal asymmetry. With a measure of asymmetry, the statistics on asymmetry for normal and pathological anatomical structures can be compared. Symmetry is a well-known mathematical group theoretical concept. In this paper, we will mathematically define the concept of weak symmetry, including topological symmetry, which serves as a basis for quantizing asymmetry. The methodology is based on non-rigid registration in the sense that the "size" of a diffeomorphism describes the amount of asymmetry. We will define this size in terms of the minimum biological work needed. That is, we evaluate how much work the biological system must carry out in order to make the object symmetrical; or identically, how much work has been carried out in order to make the ideal symmetrical object into the current (slightly) asymmetrical object. The quantization of asymmetry is validated on a set of normal (assumed near symmetrical) mandibles, and a set of pathological assumed non-symmetric mandibles exhibiting a statistically significant increase of asymmetry.

Shape analysis of hippocampi in schizophrenia has been preformed previously using the spherical harmonic SPHARM description. In these studies, the left and right hippocampi are aligned independently and the spatial relation between them is not explored. This paper presents a new SPHARM-based technique which examines not only the individual shape information of the two hippocampi but also the spatial relation between them. The left and right hippocampi are treated as a single shape configuration. A ploy-shape alignment algorithm is developed for aligning configurations of multiple SPHARM surfaces as follows: (1) the total volume is normalized; (2) the parameter space is aligned for creating the surface correspondence; (3) landmarks are created by a uniform sampling of multiple surfaces for each configuration; (4) a quaternion-based algorithm is employed to align each landmark representation to the mean configuration through the least square rotation and translation iteratively until the mean converges. After applying the poly-shape alignment algorithm, a point distribution model is applied to aligned landmarks for feature extraction. Classification is performed using Fisher's linear discriminant with an effective feature selection scheme. Applying the above procedure to our hippocampal data (14 controls versus 25 schizophrenics, all right-handed males), we achieve the best cross-validation accuracy of 92%, supporting the idea that the whole shape configuration of the two hippocampi provides valuable information in detecting schizophrenia. The results of an ROC analysis and a visualization of discriminative patterns are also included.

Image registration is a very important procedure in medical imaging analysis. However, the intensive computations involved in image registration have to some extent made it impractical for interactive use as well as limiting its general availability. This paper presents our current Grid project to facilitate image registration tasks. We have set up an image registration Grid by combining the attractive features of both Globus and Condor distributed computing environments. In order to make it more convenient to use, we have also developed a web interface for potential clients to specify and submit their image registration jobs to the Grid. The initial experiments in 3D breast MR images have shown encouraging results and demonstrated the suitability of Grid technology to this type of computationally intensive applications. The image registration Grid makes it much more straightforward for different institutes to use the identical registration program and protocols to register images consistently, quickly and efficiently. This can greatly improve data sharing and comparative studies in multi-centre trials. The Grid presented here could be an important step for clinical applications of image registration. Future work will focus on refining the Grid with the aim of upgrading it to a Grid Service and testing the system more extensively with medical imaging dataset.

B-splines are a well-known approach for non-rigid image registration. Though successfully applied to various medical applications they exhibit a high computational complexity mainly because of the lack of dedicated optimization methods. In this work we focus on a Levenberg-Marquardt type optimization routine. As a similarity measure we use least-squares functionals such as the sum of squared differences, the cross-correlation and the local correlation measure, respectively. The latter is used for multi-modality registration tasks. The proposed registration algorithm consists of three main parts. In each iteration step one has to (a) build a linear system of equations, (b) solve this system and compute an update, (c) determine the step length for the following iteration step. Appropriate stopping criteria ensure the termination of the registration task. A standard approach for (c) and several modifications are investigated. Using a quadratic model we are able to avoid additional execution of (b) during the step length adaption. Several solvers (Cholesky, CG, pre-conditioning) for (b) have been evaluated. Also, modifications on the most time consuming task (a) are investigated, leading to a speed-up by a factor up to 30. Finally, the algorithm is embedded in a multi-scale framework (both on image and on parameter level) providing additional regularization, an increased capture range and speed-up. Convergence tests have been successfully applied for a priori known transformations. Feasibility of the proposed approach is also shown for clinical applications including PET-CT registrations (19 data sets) and MR mammography.

Automated registration of multiple mammograms for CAD depends on accurate nipple identification. We developed two new image analysis techniques based on geometric and texture convergence analyses to improve the performance of our previously developed nipple identification method. A gradient-based algorithm is used to automatically track the breast boundary. The nipple search region along the boundary is then defined by geometric convergence analysis of the breast shape. Three nipple candidates are identified by detecting the changes along the gray level profiles inside and outside the boundary and the changes in the boundary direction. A texture orientation-field analysis method is developed to estimate the fourth nipple candidate based on the convergence of the tissue texture pattern towards the nipple. The final nipple location is determined from the four nipple candidates by a confidence analysis. Our training and test data sets consisted of 419 and 368 randomly selected mammograms, respectively. The nipple location identified on each image by an experienced radiologist was used as the ground truth. For 118 of the training and 70 of the test images, the radiologist could not positively identify the nipple, but provided an estimate of its location. These were referred to as invisible nipple images. In the training data set, 89.37% (269/301) of the visible nipples and 81.36% (96/118) of the invisible nipples could be detected within 1 cm of the truth. In the test data set, 92.28% (275/298) of the visible nipples and 67.14% (47/70) of the invisible nipples were identified within 1 cm of the truth. In comparison, our previous nipple identification method without using the two convergence analysis techniques detected 82.39% (248/301), 77.12% (91/118), 89.93% (268/298) and 54.29% (38/70) of the nipples within 1 cm of the truth for the visible and invisible nipples in the training and test sets, respectively. The results indicate that the nipple on mammograms can be detected accurately. This will be an important step towards automatic multiple image analysis for CAD techniques.

We present a novel in-plane rotation invariant cost function for 2D/3D registration utilizing projection-invariant transformation properties and the decomposition of the X-ray nad the DRR under comparision into orhogonal Zernike moments. As a result, only five dof have to be optimized, and the number of iteration necessary for registration can be significantly reduced. Results in a phantom study show that an accuracy of approximately 0.7° and 2 mm can be achieved using this method. We conclude that reduction of coupled dof and usage of linear independent coefficients for cost function evaluation provide intersting new perspectives for the field of 2D/3D registration.

We present a super-fast and parameter-free algorithm for non-rigid elastic registration of images of a serially sectioned whole rat brain. The purpose is to produce a three-dimensional high-resolution reconstruction. The registration is modelled as a minimization problem of a functional consisting of a distance measure and a regularizer based on the elastic potential of the displacement field. The minimization of the functional leads to a system of non-linear partial differential equations, the so-called Navier-Lame equations (NLE). Discretization of the NLE and a fixed point type iteration method lead to a linear system of equations, which has to be solved at each iteration step. We not only present a super-fast solution technique for this system, but also come up with sound strategies for accelerating the outer iteration. This does include a multi-scale approach based on a Gaussian pyramid as well as a clever estimation of the material constants for the elastic potential. The results of the registration process were controlled by an expert who was able to recognize histological details like laminations which was not possible before. Therefore, it is essential to apply elastic registration to this kind of imaging problem. Finally, the visually pleasing results were quantified by a distance measure leading to an improvement of about 79% after just 35 iteration steps.

Science and Technology International (STI) presents a novel multi-modal elastic image registration approach for a new hyperspectral medical imaging modality. STI's HyperSpectral Diagnostic Imaging (HSDI) cervical instrument is used for the early detection of uterine cervical cancer. A Computer-Aided-Diagnostic (CAD) system is being developed to aid the physician with the diagnosis of pre-cancerous and cancerous tissue regions. The CAD system uses the fusion of multiple data sources to optimize its performance. The key enabling technology for the data fusion is image registration. The difficulty lies in the image registration of fluorescence and reflectance hyperspectral data due to the occurrence of soft tissue movement and the limited resemblance of these types of imagery. The presented approach is based on embedding a reflectance image in the fluorescence hyperspectral imagery. Having a reflectance image in both data sets resolves the resemblance problem and thereby enables the use of elastic image registration algorithms required to compensate for soft tissue movements. Several methods of embedding the reflectance image in the fluorescence hyperspectral imagery are described. Initial experiments with human subject data are presented where a reflectance image is embedded in the fluorescence hyperspectral imagery.

The determination of the treatment parameters in radiation therapy requires the segmentation of the patient anatomy. This procedure is usually performed by manual contouring of 2D slices, which may require several hours. The burden is considerably increased in the context of IMRT and 4D adaptive radiotherapy. Intensity Modulated Radiation Therapy (IMRT) offers the increased ability to accurately conform the radiation dose to the target. IMRT is typically applied in fractions over 30 days. 4D or adaptive radiotherapy aims at compensating the significant anatomical changes during the course of treatment based on additional imagery. The development of fast and robust automated segmentation tools is crucial for these novel treatment methods to succeed. The purpose of this paper is to automate organ contouring of 3D CT data and time series of 3D CT in radiation therapy planning (RTP). Automated organ delineation in CT is challenging due to poor soft tissue contrast and high inter- and intra-patient organ variability. This paper presents an automated model-based concept for organ delineation, based on adaptation of 3D deformable surface models to the boundaries of the anatomical structures of interest in 3D CT and time-series thereof. A feasibility study with 40 3D clinical datasets and a 3D time series with 19 datasets was done for the risk organs (bladder, rectum, and femoral heads) of the pelvic area. The results of the validation study show that the presented model-based approach is accurate (1-1.7 mm mean error) for the tested anatomical structures, and allows a significant reduction of time compared to manual organ contouring (minutes vs. hours).

The class of deformable models has been broadly used in the segmentation of medical images. We propose a fast linkage contour, which is guided by some simple evolution rules, to extract the boundaries of objects in 2D images. A Moving Linkage consists of links, each of which is further composed of a tail joint and a head joint. Classified into horizontal, vertical and frozen (unmoved) types, the links are constricted on grids by the horizontal and vertical tracks corresponding to the image pixels. Drawing an analogy between an electric current flowing along a circuit line in the presence of magnetic fields and how it reacts in response to the force, we use the right-hand rule to guide the motion of each link. In our approach, a small closed linkage contour is initialized inside the region of interest. Stealing the concept of entropy conditions from level sets, the linkage contour is easily grown, split and merged governed by some simple deformation rules. The frozen link occurs once the speed of the free link (horizontal or vertical) is less than a specified threshold due to significant image gradients. The deformation procedure is repeated until all links become frozen, when the linkage contour resides on the edge of the segmented object. This new deformable model was validated by the preliminary segmentation results regarding phantom and medical images.

Object segmentation is of paramount interest in many medical imaging applications. Among others, "snake"-an "active contour"-is a popular boundary-based segmentation framework where a spline is continuously deformed to lock onto an object boundary. The dynamics of a snake is governed by different internal and external forces. A major limitation of this framework has been the difficulty in using object-intensity driven features into snake dynamics which may prevent uncontrolled expansion/contraction once the snake leaks through a weak boundary region. In this paper, object-intensity force is effectively introduced into the snake-model using class uncertainty theory. Given a priori knowledge of object/background intensity distributions, class uncertainty yields object/background class of any location and establishes the confidence level of the classification. This confidence level has previously been demonstrated to be high inside the object/background regions and low near boundaries with intermediate intensities. This class uncertainty information adds an expanding (outward) force at locations pertaining to intensity-based object class and a squeezing (inward) force inside background regions. Consequently, the method possesses potential to resist an uncontrolled expansion of the snake (for an expanding type) into the background through a weak boundary while reducing the effect of this force near the boundary using the confidence information. The theory of object class uncertainty induced snake is developed and an implementation with efficient graphical interface is achieved. Preliminary results of application of the proposed snake approach on different images are presented and comparisons with conventional snake approaches are demonstrated.

In order to complete a thorough examination of a patient heart muscle, physicians practice two common invasive procedures: the ventriculography, which allows the determination of the ejection fraction, and the coronarography, giving among other things, information on stenosis of arteries. We propose a method that allows the determination of a contraction index similar to ejection fraction, using only single-plane coronarography. Our method first reconstructs in 3D, selected points on the angiogram, using a 3D model devised from data published by Dodge ea. ['88, '92]. We then follow the point displacements through a complete heart contraction cycle. The objective function, minimizing the RMS distances between the angiogram and the model, relies on affine transformations, i.e. translation, rotation and isotropic scaling. We validate our method on simulated projections using cases from Dodge data. In order to avoid any bias, a leave-one-out strategy was used, which excludes the reference case when constructing the 3D coronary heart model. The simulated projections are created by transforming the reference case, with scaling, translation and rotation transformations, and by adding random 3D noise for each frame in the contraction cycle. Comparing the true scaling parameters to the reconstructed sequence, our method is quite robust (R²=96.6%, P<1%), even when noise error level is as high as 1 cm. Using 10 clinical cases we then proceeded to reconstruct the contraction sequence for a complete cardiac cycle starting at end-diastole. A simple heart contraction mathematical model permitted us to link the measured ejection fraction of the different cases to the maximum heart contraction amplitude (R²=57%, P<1%) determined by our method.

Automatic retrospective motion correction algorithms based on iterative optimization of an image quality measure have been demonstrated in a variety of MRI acquisitions. These algorithms are computationally intensive and may require several minutes per image or more. One computational bottleneck is the need for an inverse FFT at each iteration to reconstruct and evaluate the image. We describe a method for performing the iterative search primarily in image space, greatly reducing the number of FFTs required. This can significantly increase the computational speed, particularly when the evaluation is performed only on a sub-region of the image.

Accurate, robust, and efficient automated segmentation of anatomical structures is difficult because of their complexity and inter-patient and inter-acquisition variations in shape, appearance, and spatial arrangement. To deal with the complexity and variations of anatomical structures, a number of model-based segmentation methods (e.g. snakes, shape and appearance models, image templates, etc.) have been proposed in the past. However, because applying individual models often does not lead to robust and accurate segmentation, we propose a framework for image segmentation by connected parametrical models. In our framework, which is a generalization of Fischler and Elschlager's pictorial structure (IEEE TC, 1973) and Bernard's et al. hierarchical principal component analysis framework (MICCAI, 2001), models cooperate via their parameters at different levels of hierarchy. In parametrical space, defined by parameters of all cooperating models, parameters may be connected by physical or statistical models. Connections may work either between intrinsic parameters of individual models, such as position, orientation and scale, or alternatively between arbitrarily transformed parameters, obtained, for instance, by PCA. By selecting appropriate parametrical models and models of connections between them, segmentation by the proposed framework has the potential to achieve accurate and robust results on a large variety of medical images. A framework, incorporating physical and/or statistical models of connections between intrinsic or transformed parameters is demonstrated on real images.

Radiologists perform a CT Angiography procedure to examine vascular structures and associated pathologies such as aneurysms. Volume rendering is used to exploit volumetric capabilities of CT that provides complete interactive 3-D visualization. However, bone forms an occluding structure and must be segmented out. The anatomical complexity of the head creates a major challenge in the segmentation of bone and vessel. An analysis of the head volume reveals varying spatial relationships between vessel and bone that can be separated into three sub-volumes: “proximal”, “middle”, and “distal”. The “proximal” and “distal” sub-volumes contain good spatial separation between bone and vessel (carotid referenced here). Bone and vessel appear contiguous in the “middle” partition that remains the most challenging region for segmentation. The partition algorithm is used to automatically identify these partition locations so that different segmentation methods can be developed for each sub-volume. The partition locations are computed using bone, image entropy, and sinus profiles along with a rule-based method. The algorithm is validated on 21 cases (varying volume sizes, resolution, clinical sites, pathologies) using ground truth identified visually. The algorithm is also computationally efficient, processing a 500+ slice volume in 6 seconds (an impressive 0.01 seconds / slice) that makes it an attractive algorithm for pre-processing large volumes. The partition algorithm is integrated into the segmentation workflow. Fast and simple algorithms are implemented for processing the “proximal” and “distal” partitions. Complex methods are restricted to only the “middle” partition. The partitionenabled segmentation has been successfully tested and results are shown from multiple cases.

Phase-contrast magnetic resonance angiography (PC-MRA) can produce phase images which are 3-dimensional pictures of vascular structures. However, it also provides magnitude images, containing anatomical - but no vascular - data. Classically, algorithms dedicated to PC-MRA segmentation detect the cerebral vascular tree by only working on phase images. We propose here a new approach for segmentation of cerebral blood vessels in PC-MRA using both types of images. This approach is based on the hypothesis that a magnitude image contains anatomical information useful for vascular structures detection. That information can then be transposed from a normal case to any patient image by image registration. An atlas of the whole head has been developed in order to store such anatomical knowledge. It divides a magnitude image into several "vascular areas", each one having specific vessel properties. The atlas can be applied on any magnitude image of an entire or nearly entire head by deformable matching, thus helping to segment blood vessels from the associated phase image. The segmentation method used afterwards is composed of a topology-conserving region growing algorithm using adaptative threshold values depending on the current region of the atlas. This algorithm builds the arterial and venous trees by iteratively adding voxels which are selected according to their greyscale value and the variation of values in their neighborhood. The topology conservation is guaranteed by only selecting simple points during the growing process. The method has been performed on 15 PC-MRA's of the brain. The results have been validated using MIP and 3D surface rendering visualization; a comparison to other results obtained without an atlas proves that atlas-based methods are an effective way to optimize vascular segmentation strategies.

The quantitative analysis of cardiac cine MRI sequences requires automated, robust, and fast image processing algorithms for the 4D (3D + time) segmentation of the heart chambers. The use of shape models has proven efficient in extracting the cardiac volumes for single phases, but less attention has been focused on incorporating prior knowledge about the cardiac motion. To explicitly address the temporal aspect of the segmentation problem, this paper proposes a full Bayesian model, where the prior information is represented by a cardiac shape and motion model. In this framework, the solution of the segmentation is defined by means of a probability distribution over the parameters of the space-time problem. The computed solution, obtained by means of sequential Monte Carlo techniques, has the advantage of being both spatially and temporally coherent. Furthermore, the method does not require any particular representation of the shape or of the motion model; it is therefore generic and highly flexible.

This paper presents a method for segmenting internal brain structures in MR images. It introduces prior information in an original way through descriptions of the spatial arrangement of structures by means of spatial relations, which are represented in the fuzzy set framework. The method is hierarchical as the segmentation of a given structure is based on the previously segmented ones. The processing of each structure is decomposed into two stages: an initialization stage which makes extensive use of prior knowledge and a refinement stage using a 3D deformable model. The deformable model is guided by an external force representing the combination of a classical data term derived from an edge map and a force corresponding to a given spatial relation. We propose different ways to compute a force from a fuzzy set representing a relation or a combination of relations. Results obtained for the lateral ventricles, the third ventricle, the caudate nuclei and the thalami are promising. The proposed combination of spatial relations and deformable models has proved to be very useful to segment parts of the structures were no visible edges are present, improving the segmentation accuracy.

Computed Tomography Angiography (CTA) is an emerging modality for assessing cardiac anatomy. The delineation of the cardiac volume of interest (VOI) is a pre-processing step for subsequent visualization or image processing. It serves the suppression of anatomic structures being not in the primary focus of the cardiac application, such as sternum, ribs, spinal column, descending aorta and pulmonary vasculature. These structures obliterate standard visualizations such as direct volume renderings or maximum intensity projections. In addition, outcome and performance of post-processing steps such as ventricle suppression, coronary artery segmentation or the detection of short and long axes of the heart can be improved. The structures being part of the cardiac VOI (coronary arteries and veins, myocardium, ventricles and atria) differ tremendously in appearance. In addition, there is no clear image feature associated with the contour (or better cut-surface) distinguishing between cardiac VOI and surrounding tissue making the automatic delineation of the cardiac VOI a difficult task. The presented approach locates in a first step chest wall and descending aorta in all image slices giving a rough estimate of the location of the heart. In a second step, a Fourier based active contour approach delineates slice-wise the border of the cardiac VOI. The algorithm has been evaluated on 41 multi-slice CT data-sets including cases with coronary stents and venous and arterial bypasses. The typical processing time amounts to 5-10s on a 1GHz P3 PC.

Segmentation of colon CT images is the main factor that inhibits automation of virtual colonoscopy. There are two main reasons that make efficient colon segmentation difficult. First, besides the colon, the small bowel, lungs, and stomach are also gas-filled organs in the abdomen. Second, peristalsis or residual feces often obstruct the colon, so that it consists of multiple gas-filled segments. In virtual colonoscopy, it is very useful to automatically connect the centerlines of these segments into a single colon centerline. Unfortunately, in some cases this is a difficult task. In this study a novel method for automated colon segmentation and connection of colon segments' centerlines is proposed. The method successfully combines features of segments, such as centerline and thickness, with information on main colon segments. The results on twenty colon cases show that the method performs well in cases of small obstructions of the colon. Larger obstructions are mostly also resolved properly, especially if they do not appear in the sigmoid part of the colon. Obstructions in the sigmoid part of the colon sometimes cause improper classification of the small bowel segments. If a segment is too small, it is classified as the small bowel segment. However, such misclassifications have little impact on colon analysis.

Tumor definition and diagnosis require the analysis of the spatial distribution and Hounsfield unit (HU) values of voxels in computed tomography (CT) images, coupled with a knowledge of normal anatomy. Segmentation of the tumor in neuroblastoma is complicated by the fact that the mass is almost always heterogeneous in nature; furthermore, viable tumor, necrosis, fibrosis, and normal tissue are often intermixed. Rather than attempt to separate these tissue types into distinct regions, we propose to explore methods to delineate the normal structures expected in abdominal CT images, remove them from further consideration, and examine the remaining parts of the images for the tumor mass. We explore the use of fuzzy connectivity for this purpose. Expert knowledge provided by the radiologist in the form of the expected structures and their shapes, HU values, and radiological characteristics are also incorporated in the segmentation algorithm. Segmentation and analysis of the tissue composition of the tumor can assist in quantitative assessment of the response to chemotherapy and in the planning of delayed surgery for resection of the tumor. The performance of the algorithm is evaluated using cases acquired from the Alberta Children's Hospital.

This paper presents a novel segmentation technique to extract cavity contours from ultrasound images. The problem is first discretized by projecting equispaced radii from an arbitrary seed point inside the cavity towards its boundary. The distance of the cavity boundary from the seed point is modeled by the trajectory of a moving object. The motion of this moving object is assumed to be governed by a finite set of dynamical models subject to uncertainty. Candidate edge points obtained along each radius may include the actual object position and some false measurements. This modeling approach enables us to employ the interacting multiple model (IMM) estimator along with a probabilistic data association filter (PDAF) for contour extraction. Since the method does not employ any numerical optimization, convergence is very fast. The stability and accuracy of the method is demonstrated by segmenting contours from a series of ultrasound images with different levels of speckles and echo dropouts in the cavity contour. An application of the method in segmenting bone contours from CT images is also presented. We have demonstrated that in certain situations, such as wrist CT images, when the distance between two bones is very small and conventional thresholding techniques fail, our algorithm is able to segment bones successfully.

Deformable models are important and popular techniques for extracting the shape of objects in medical images. We used the simulation of a physical system (a Charged Fluid) to guide the evolution of a propagating interface to segment objects in brain MR and CT images. The Charged Fluid was simulated as a system of charged particles that exert a repelling electric force upon each other. In our approach, the boundary of the segmenting object was determined by the image gradient, which was modeled as potential wells that stopped the propagating front. The simulation was evolved in two steps that were governed by Poisson's equation. One allowed the propagating interface of the Charged Fluid to deform into a new shape in response to the electric potential and the image potential. The other allowed Charged Fluid elements to flow along the propagating interface, which was treated as the surface of an isolated conductor, until an electrostatic equilibrium was achieved. The electric potential of the simulated system was rapidly computed from Poisson's equation by the aid of the Fast Fourier Transform. The procedure was repeated until the propagating front resided on the boundary of objects being segmented. This method was used to extract the contour of the brain surface for use in skull stripping applications.

An extension to level set based segmentation is proposed for vascular tree delineation. The method starts with topology extraction, by a shape constrained level set evolution steered by a strictly positive, image base speed function to ensure some oversegmentation. Next, the skeleton of the resulting oversegmentation is determined, which then is used to initialise another level set steered by a speed function with both negative and positive speed forces based on image features, to obtain a most accurate segmentation. The novelty of our approach lies in the shape constraint that is imposed implicitly on the first level set evolution. We apply repeatedly re-initializations of this evolution with a topology preserving skeleton of the current zero level set. We compare this method with a plain level set evolution steered by the same full range speed function. Both are initialised by placing a single seed point at the root of the vessel tree. Pilot experiments on twelve multislice CT data sets of the Circle of Willis show that our method is capable of segmenting the smaller branches at the distal part of the vessel tree structures and has the potential to segment vessels which are distal to a severe stenosis or occlusion.

Intensity differences between objects are often used for segmentation. In an ideal situation, these differences permit the computation of edges that form complete contours around objects in the image. However, edges found in real images are usually a set of real and spurious disconnected boundary segments. Even more challenging are those so called apparent or subjective contours whose boundary are not defined by intensity or texture variations. In this paper, we present a novel method to segment and reconstruct images with missing boundaries, including images with large missing edges commonly found in ultrasound imaging. We test our algorithm on classic synthetic images, phantom images and on real ultrasound images of the bladder, heart, and colon.

Segmentation of three-dimensional anatomical brain images into tissue classes has applications in both clinical and research settings. This paper presents the implementation and quantitative evaluation of a four-phase three-dimensional active contour implemented with a level set framework for automated segmentation of brain MRIs. The segmentation algorithm performs an optimal partitioning of three-dimensional data based on homogeneity measures that naturally evolves to the extraction of different tissue types in the brain. Random seed initialization was used to speed up numerical computation and avoid the need for a priori information. This random initialization ensures robustness of the method to variation of user expertise, biased a priori information and errors in input information that could be influenced by variations in image quality. Experimentation on three MRI brain data sets showed that an optimal partitioning successfully labeled regions that accurately identified white matter, gray matter and cerebrospinal fluid in the ventricles. Quantitative evaluation of the segmentation was performed with comparison to manually labeled data and computed false positive and false negative assignments of voxels for the three organs. We report high accuracy for the two comparison cases. These results demonstrate the efficiency and flexibility of this segmentation framework to perform the challenging task of automatically extracting brain tissue volume contours.

This paper focuses on the problem of accurately extracting the CSF-tissue boundary, particularly around the ventricular surface, from serial structural MRI of the brain acquired in imaging studies of aging and dementia. This is a challenging problem because of the common occurrence of peri-ventricular lesions which locally alter the appearance of white matter. We examine a level set approach which evolves a four dimensional description of the ventricular surface over time. This has the advantage of allowing constraints on the contour in the temporal dimension, improving the consistency of the extracted object over time. We follow the approach proposed by Chan and Vese which is based on the Mumford and Shah model and implemented using the Osher and Sethian level set method. We have extended this to the 4 dimensional case to propagate a 4D contour toward the tissue boundaries through the evolution of a 5D implicit function. For convergence we use region-based information provided by the image rather than the gradient of the image. This is adapted to allow intensity contrast changes between time frames in the MRI sequence. Results on time sequences of 3D brain MR images are presented and discussed.

In this work we describe an application of the Support Vector Machine (SVM) classifier for the segmentation of wounds in color images. The SVM-based segmentation combines naturally a high dimensional space of image features into a single classification machine. Since particular choice of image features is crucial for the performance of SVM classifier, we investigate the efficiency of color- and texture-based features for the differentiation between skin and wound tissue. We find that color features provide better separation between these two tissues. However, incorporation of even a single textural feature improves an overall quality of the SVM classification. We test the impact of each color feature on the quality of wound segmentation and find optimal combination of these features which produces best segmentation result. We suggest a Histogram Sampling technique, which gives wider separation between wound and skin in the color space. Finally, we find a set of image features, which is typical for most types of wounds. When these features are used as an input to the SVM classifier, a fairly robust segmentation of different wound types is achieved. We evaluate the performance of SVM-based segmentation using ground-truth segmentation carried out by clinicians.

This paper describes an automatic method for measuring the biparietal diameter (BPD) and head circumference (HC) in ultrasound fetal images. A total of 217 ultrasound images were segmented by using a K-Mean classifier, and the head skull was detected in 214 of the 217 cases by an iterative randomized Hough transform developed for detection of incomplete curves in images with strong noise without user intervention. The automatic measurements were compared with conventional manual measurements by sonographers and a trained panel. The inter-run variations and differences between the automatic and conventional measurements were small compared with published inter-observer variations. The results showed that the automated measurements were as reliable as the expert measurements and more consistent. This method has great potential in clinical applications.

We present new approaches based on Genetic Algorithms (GAs), Simulated Annealing (SA) and Expectation Maximization (EM) for learning parameters of the mixture of Gaussian model. GAs are adaptive search techniques designed to search for near optimal solutions of large-scale optimization problems with multiple local maxima. It has been shown that GAs are independent of initialization parameters and can provide an efficient technique to optimize functions in large search spaces while the solution obtained by EM is a function of initial parameters, hence relatively high likelihood of achieving sub-optimal solution, due to trapping in local maxima. In this work we propose a new incorporate genetic algorithm with EM (Interactive GA-EM) to improve estimation of Gaussian mixture parameters. The method uses a population of mixture models, rather than a single mixture, interactively in both GA and EM to determine Gaussian mixture parameters. To assess the performance of the proposed methods, a series of Gaussian phantoms, based on modified Shepp-Logan method, were created. All proposed methods were employed to estimate the tissue parameters in each phantom. The results indicate that the EM algorithm, as expected, is heavily impacted by the initial values. The best result on both computational time and accuracy was obtained from Interactive GA-EM. The proposed method offers an accurate and stable solution for parameter estimation on Gaussian mixture models, with higher chance of achieving global optimal. Obtaining such accurate parameter estimation is a key requirement for several image segmentation approaches, which rely on a priori knowledge of tissue distribution.

Fuzzy C-means (FCM), in spite of its potent advantages in exploratory analyze of functional magnetic resonance imaging (fMRI), suffers from limitations such as a priori determination of number of clusters, unknown statistical significance for the results, and instability of the results when it is applied on raw fMRI time series. Choosing different number of clusters, or thresholding the membership degree at different levels, lead to considerably different activation maps. However, research work for finding a standard index to determine the number of clusters has not yet succeeded. Using randomization, we developed a method to control false positive rate in FCM, which gives a meaningful statistical significance to the results. Making use of this novel method and an ROC-based cluster validity measure, we determined the optimal number of clusters. In this study, we applied the FCM on a feature space that takes the variability of hemodynamic response function into account (HRF-based feature space). The proposed method found the accurate number of clusters in simulated fMRI data. In addition, the proposed method generated excellent results for experimental fMRI data and showed a good reproducibility for determining the number of clusters.

Architectural distortion is a subtle abnormality in mammograms, and a source of overlooking errors by radiologists. Computer-aided diagnosis (CAD) techniques can improve the performance of radiologists in detecting masses and calcifications; however, most CAD systems have not been designed to detect architectural distortion. We present a new method to detect and localize architectural distortion by analyzing the oriented texture in mammograms. A bank of Gabor filters is used to obtain the orientation field of the given mammogram. The orientation field is filtered and downsampled, to reduce noise and also to reduce the computational effort required by the subsequent methods. The downsampled orientation field is analyzed to produce three phase portrait maps: node, saddle, and spiral. The node map is linearly filtered, thresholded, and morphologically filtered to detect architectural distortion. The method was tested with 18 mammograms containing architectural distortion. In a preliminary experiment, a sensitivity of 88% was obtained at 15 false positives per image. Several possibilities for the improvement of the technique are being explored. A qualitative analysis of the performance of the method with stellate lesions indicates potential for enhancement of the technique.

Content-based image retrieval requires a formal description of visual information. In medical applications, all relevant biological objects have to be represented by this description. Although color as the primary feature has proven successful in publicly available retrieval systems of general purpose, this description is not applicable to most medical images. Additionally, it has been shown that global features characterizing the whole image do not lead to acceptable results in the medical context or that they are only suitable for specific applications. For a general purpose content-based comparison of medical images, local, i.e. regional features that are collected on multiple scales must be used. A hierarchical attributed region adjacency graph (HARAG) provides such a representation and transfers image comparison to graph matching. However, building a HARAG from an image requires a restriction in size to be computationally feasible while at the same time all visually plausible information must be preserved. For this purpose, mechanisms for the reduction of the graph size are presented. Even with a reduced graph, the problem of graph matching remains NP-complete. In this paper, the Similarity Flooding approach and Hopfield-style neural networks are adapted from the graph matching community to the needs of HARAG comparison. Based on synthetic image material build from simple geometric objects, all visually similar regions were matched accordingly showing the framework's general applicability to content-based image retrieval of medical images.

In contrast-enhanced (CE) MRI of the breast, signal-intensity time curves have been proven useful in differentiating between benign and malignant lesions. Due to uptake heterogeneity in the breast lesion, however, the signal-intensity time curve obtained from a specific region in the lesion may outperform that from the entire lesion. In this study, we propose the use of fuzzy c-means (FCM) clustering algorithms to reveal different temporal patterns within the breast lesion. The algorithm finds fuzzy cluster centers (i.e., temporal patterns) and assigns membership values to each voxel. The temporal pattern with maximum initial enhancement is selected as the representative curve of the lesion and the thresholded membership map is the identified region of fast enhancement. The approach was applied to the analysis of 121 lesions (77 malignant and 44 benign). The resulting representative curves were classified with linear discriminant analysis (LDA). The differentiation performance of LDA output in leave-one-out cross evaluation was assessed using receiver operating characteristic (ROC) analysis. Our results show that the use of FCM significantly improved the performance of signal-intensity time curves in the task of distinguishing between malignant and benign lesions.

This paper describes a method for automatically analysing and segmenting the corpus callosum from magnetic resonance images of the brain based on the widely used Active Appearance Models (AAMs) by Cootes et al. Extensions of the original method, which are designed to improve this specific case are proposed, but all remain applicable to other domain problems. The well-known multi-resolution AAM optimisation is extended to include sequential relaxations on texture resolution, model coverage and model parameter constraints. Fully unsupervised analysis is obtained by exploiting model parameter convergence limits and a maximum likelihood estimate of shape and pose. Further, the important problem of modelling object neighbourhood is addressed. Finally, we describe how correspondence across images is achieved by selecting the minimum description length (MDL) landmarks from a set of training boundaries using the recently proposed method of Davies et al. This MDL-approach ensures a unique parameterisation of corpus callosum contour variation, which is crucial for neurological studies that compare reference areas such as rostrum, splenium, et cetera. We present quantitative and qualitative results that show that the method produces accurate, robust and rapid segmentations in a cross sectional study of 17 subjects, establishing its feasibility as a fully automated clinical tool for analysis and segmentation.

In this paper, a novel polynomial-time algorithm is described for solving the optimal net surface detection problem on proper ordered multi-column graphs in N-D space (N ≥ 3). The method is applied to searching for optimal object boundaries with arbitrary smoothness constraints in volumetric medical images. By simple transformations, such optimal surface detection problems can be simplified to a problem of computing the minimum s-t cuts in the transformed graphs. An efficient implementation for the 3-D case that can achieve near real-time performance on moderate-sized datasets is presented. We further examine our technique in experiments by segmenting the cylindrical surfaces of human airways from pulmonary volumetric CT images, and compare the results to those produced by previous methods. By allowing full specifications of the cost-function and smoothness constraints without degrading the performance, the new algorithm is more flexible than traditional methods and guarantees global optimality. The multi-dimensional nature of the algorithm maintains continuity in higher dimensions.

This paper presents a generic and unified method to identify a set of anatomical landmarks of interest within the medical image domains. Landmark identification is important as it provides us with: 1) initial information for registration, 2) navigation and retrieval guidance through the image data, 3) initial models for segmentation, and 4) valuable (though rough) information about the organs/structures of interest. The proposed method initially uses a supervised learning procedure and then improves itself based on the Bayes’ theory. The procedure at the first step requires an expert to define a rough roadmap passing through a set of high-contrast landmarks (milestones), and eventually reaching at the structure of interest. The expert is asked to mark the milestones as desired points and a few points around them as undesired points, respectively. Then we estimate Gaussian models for the marked points by which the optimal search area for each desired landmark is determined. The search areas estimated at this step are considered as the segments of the statistical roadmap. An additional set of statistical models along with the above ones are used to form a set of rules to evaluate the points being found during the search procedure. The points that satisfy the rules will be recognized as the landmarks of interest. As the above method is being applied on a set of new patients/cases, a set of valid landmarks of interests becomes available. This new piece of information is then being used to modify the current statistical roadmap based on the Bayes' theory. We have applied the proposed method on T1-weighted brain MRI of 10 epileptic patients to find the landmarks of the hippocampus. In our experiment, six patients formed the training set, and we observed one-step iteration of the Bayesian modification. The method made no false alarms. The overall success rate (average of sensitivity and specificity) of the algorithm was 83.3% with an accuracy of 99.2%. In localizing the hippocampus, the proposed method (with almost perfect results) was 600 times faster than the mutual information registration (with poor and partly wrong results).

In 2002, at this meeting, we presented a framework for the evaluation of image segmentation methods. That framework suggested the consideration of both recognition and delineation tasks and described three groups of factors under precision, accuracy, and efficiency for evaluating segmentation methods. In this paper, focusing on accuracy, we argue that evaluations conducted on particular manifestations of any segmentation method M (as determined by fixed values of the parameters of M) do not characterize the range of behavior of M. To capture the entire range of behavior of M for a given application domain, we describe a method called Delineation Operating Characteristic (DOC) analysis. The DOC curve summarizes the behavior of M and the tradeoff that is inherent in M between false positive and false negative regions within the regions delineated by M. This, we argue, will allow more complete understanding and more meaningful comparison of the relative accuracy of different methods. This analysis may also allow the selection of more optimal manifestations of M for a given application domain. By way of illustration, we demonstrate with examples how DOC curves can be used for comparing the accuracy of several segmentation methods.

In this work we compare the performance of a number of vessel segmentation algorithms on a newly constructed retinal vessel image database. Retinal vessel segmentation is important for the detection of numerous eye diseases and plays an important role in automatic retinal disease screening systems. A large number of methods for retinal vessel segmentation have been published, yet an evaluation of these methods on a common database of screening images has not been performed. To compare the performance of retinal vessel segmentation methods we have constructed a large database of retinal images. The database contains forty images in which the vessel trees have been manually segmented. For twenty of those forty images a second independent manual segmentation is available. This allows for a comparison between the performance of automatic methods and the performance of a human observer. The database is available to the research community. Interested researchers are encouraged to upload their segmentation results to our website (http://www.isi.uu.nl/Research/Databases). The performance of five different algorithms has been compared. Four of these methods have been implemented as described in the literature. The fifth pixel classification based method was developed specifically for the segmentation of retinal vessels and is the only supervised method in this test. We define the segmentation accuracy with respect to our gold standard as the performance measure. Results show that the pixel classification method performs best, but the second observer still performs significantly better.

Enhancements are described for an approach that unifies edge preserving smoothing with segmentation of time sequences of volumetric images, based on differential edge detection at multiple spatial and temporal scales. Potential applications of these 4-D methods include segmentation of respiratory gated positron emission tomography (PET) transmission images to improve accuracy of attenuation correction for imaging heart and lung lesions, and segmentation of dynamic cardiac single photon emission computed tomography (SPECT) images to facilitate unbiased estimation of time-activity curves and kinetic parameters for left ventricular volumes of interest. Improved segmentation of lung surfaces in simulated respiratory gated cardiac PET transmission images is achieved with a 4-D edge detection operator composed of edge preserving 1-D operators applied in various spatial and temporal directions. Smoothing along the axis of a 1-D operator is driven by structure separation seen in the scale-space fingerprint, rather than by image contrast. Spurious noise structures are reduced with use of small-scale isotropic smoothing in directions transverse to the 1-D operator axis. Analytic expressions are obtained for directional derivatives of the smoothed, edge preserved image, and the expressions are used to compose a 4-D operator that detects edges as zero-crossings in the second derivative in the direction of the image intensity gradient. Additional improvement in segmentation is anticipated with use of multiscale transversely isotropic smoothing and a novel interpolation method that improves the behavior of the directional derivatives. The interpolation method is demonstrated on a simulated 1-D edge and incorporation of the method into the 4-D algorithm is described.

Related to the demand for fast and efficient tomographic reconstruction methods, the interest for Direct Fourier (DF) methods, which have a reduced computational complexity, has been growing. In this paper we present a new NUFFT-based DF reconstruction method which can be directly applied to fan-beam CT data sets avoiding the interpolation in Radon space as well as the interpolation in Fourier space. The performance of the new algorithm, in ideal and noisy conditions, is compared to those of other well known reconstruction methods, revealing an excellent behavior, specially in noisy conditions.

We propose a generalized algorithm for image reconstruction from data acquired in an asymmetric fan-beam configuration with a spatially varying focal length. This algorithm is an extension of our previous algorithm for fan-beam computed tomography (CT) that involves a shift-variant filtering. As compared to the existing algorithms, the proposed algorithm involves no additional explicit interpolation and eliminates the spatially-variant weighting factor. Therefore, it retains the spatial resolution and is less susceptible to data noise. We performed computer-simulation studies to validate and evaluate the proposed algorithm, and quantitative results demonstrate that the proposed algorithm possesses desirable noise and numerical properties.

An important breakthrough in helical cone-beam reconstruction is the development of Katsevich's algorithms, which appear to have numerous advantages over the algorithms developed previously. The original algorithm proposed by Katsevich has a simple form and requires only once the computation of the data filtering. It, however, invokes a derivative of the data function with respect to the rotation angle along the helical trajectory and thus makes the algorithm numerically susceptible to the sample aliasing along the helical trajectory. A modified version of the algorithm later developed by Katsevich that avoids the explicit computation of the derivative of the data function along the helical trajectory. However, this modified algorithm contains more terms than does the original algorithm and involves two different filtering of the data function. Therefore, this modified algorithm is computationally more complex and demanding than does the original Katsevich's algorithm. In this work, based upon the original Katsevich's algorithm, we present a new algorithm that not only avoids explicit computation of the derivative of the data function along the helical trajectory but also requires only one filtering of the data function. Therefore, in general, our algorithm is quantitatively more accurate than the original Katsevich's algorithm and is computationally more efficient than the modified Katsevich's algorithm.

With the introduction of ultra-fast cone beam scanners, cardiac CT imaging has become feasible. In order to achieve excellent image quality, cardiac phases must be found during which the heart is quasi-stationary. Electrocardiogram (ECG) information does not always correspond to the exact motion-state of the heart, and there is high patient variability with respect to the motion pattern. The clinician has to select stable phases manually without an exact knowledge about the patient-specific motion. Therefore, several high-resolution volumes corresponding to different phases have to be reconstructed, which is an inefficient task. In this contribution, a simple and efficient image-based technique is introduced which is able to deliver patient-specific stable cardiac phases in an automatic fashion. For this purpose, a low-resolution 4D data set is reconstructed in advance. The most stable phases are derived from this 4D data set by calculating the similarity between subsequent positions in the cardiac cycle. Information about the patient-specific motion of the heart can be determined. High-resolution reconstructions are shown at the automatically predicted phase points corresponding to systole and diastole. The images are superior to images reconstructed at other phase points.

With the evolution from multi-detector-row CT to cone beam volumetric CT, to maintain reconstruction accuracy becomes more challenging. To combat the severe artifacts caused by a large cone angle, three-dimensional reconstruction algorithms have to be used in CB volumetric CT. In practice, the filtered backprojection (FBP) reconstruction algorithm is the most desirable due to its pipe-line computational structure and image generation efficiency. The first CB-FBP reconstruction algorithm is the well-known FDK algorithm that was originally derived for a circular x-ray source trajectory by heuristically extending its two-dimensional counterpart. Later on, a general CB-FBP reconstruction algorithm was derived for non-circular, such as helical, source trajectories. It has been recognized that, a filtering operation in the projection data along the tangential direction of a helical x-ray source trajectory can significantly improve the reconstruction accuracy of helical CB volumetric CT. However, the tangential filtering encounters latitudinal data truncation, resulting in degraded noise characteristics or data manipulation inefficiency. A CB-FBP reconstruction algorithm using one-dimensional rotational filtering across detector rows (namely CB-RFBP) is proposed in this manuscript. In principal, the proposed CB-RFBP reconstruction algorithm is approximate, but approaches the reconstruction accuracy that can be achieved by exact helical CB-FBP reconstruction algorithms because of the rotational filtering. Unlike exact CB-FBP reconstruction algorithms in which all redundant data are usually discarded to maintain data redundancy, the proposed CB-RFBP reconstruction algorithm makes use of all available projection data, resulting in significantly improved noise characteristics and dose efficiency. Moreover, the rotational filtering across detector rows not only survives the so-called long object problem, but also avoids latitudinal data truncation existing in other helical CB-FBP reconstruction algorithm in which a tangential filtering is carried out, providing improved noise characteristics, dose efficiency or data manipulation efficiency.

In this paper, a fast, accurate and memory-saving Tomosynthesis algorithm is presented based on the Algebraic Reconstruction Technique (ART). In this approach, a one step ART iterative reconstruction takes the place of the commonly used two step Tomosynthesis reconstruction and deblurring processes. The weight matrix required by ART is calculated offline and saved in a look-up-table since the weight matrix will not change with the object if the acquisition geometries of the projections are fixed. This look-up-table speeds up the reconstruction procedure and the memory space is greatly reduced by using a compact weight matrix. A Bessel-Kaiser function is utilized in this algorithm as the pixel basis function, which improves the quality of the reconstruction over other commonly used basis functions. Simulation results show that the presented algorithm generates fast, accurate and memory-saving reconstructions of a three-dimensional object.

Interest in radial MRI, also known as projection reconstruction (PR) MRI, has increased recently for uses such as fast scanning and undersampled acquisitions. Additionally, PR acquisitions have intrinsic advantages over standard two-dimensional Fourier transform (2DFT) imaging with respect to motion of the imaged object. It is well known that alignment of each spatial domain projection's center of mass (calculated using each projection's 0th and 1st moments) to the center of the field of view corrects shifts caused by in-plane translation. Here we report a previously unrealized ability to determine the in-plane rotational motion of an imaged object using the 2nd moments of the translation-corrected spatial domain projections. The correction requires only the PR data itself and a new projection view angle acquisition time order. The proposed view angle time order allows the acquisition to be "self-navigating" with respect to both in-plane translation and rotation. Image reconstruction using the aligned projections and detected acquisition angles can eliminate or significantly reduce image artifacts due to such motion. We describe the theory of the correction technique and demonstrate its effectiveness using a computer-controlled motion phantom executing 2-D in-plane translations and a customized pulse sequence capable of introducing known in-plane rotational errors.

We present an algorithm for blindly recovering constituent source spectra from magnetic resonance spectroscopic imaging (MRSI) of human brain. The algorithm is based on the non-negative matrix factorization (NMF) algorithm, extending it to include a constraint on the positivity of the amplitudes of the recovered spectra and mixing matrices. This positivity constraint enables recovery of physically meaningful spectra even in the presence of noise that causes a significant number of the observation amplitudes to be negative. The algorithm, which we call constrained non-negative matrix factorization (cNMF), does not enforce independence or sparsity, though it recovers sparse sources quite well. It can be viewed as a maximum likelihood approach for finding basis vectors in a bounded subspace. In this case the optimal basis vectors are the ones that envelope the observed data with a minimum deviation from the boundaries. We incorporate the cNMF algorithm into a hierarchical decomposition framework, showing that it can be used to recover tissue-specific spectra, e.g., spectra indicative of malignant tumor. We demonstrate the hierarchical procedure on ¹H long echo time (TE) brain absorption spectra and conclude that the computational efficiency of the cNMF algorithm makes it well-suited for use in diagnostic work-up.

A novel approach for correcting intensity nonuniformity in magnetic resonance imaging (MRI) is presented. This approach is based on the fusion of spatial and gray-level histogram information. Spatial information about intensity nonuniformity is obtained using cubic B-spline smoothing. Gray-level histogram information of the image corrupted by intensity nonuniformity is exploited from a frequential point of view. The proposed correction method is illustrated using both physical phantom and human brain images. The results are consistent with theoretical prediction, and demonstrate a new way of dealing with intensity nonuniformity problems.

MRI data analysis is routinely done on the magnitude part of complex images. While both real and imaginary image channels contain Gaussian noise, magnitude MRI data are characterized by Rice distribution. However, conventional filtering methods often assume image noise to be zero mean and Gaussian distributed. Estimation of an underlying image using magnitude data produces biased result. The bias may lead to significant image errors, especially in areas of low signal-to-noise ratio (SNR). The incorporation of the Rice PDF into a noise filtering procedure can significantly complicate the method both algorithmically and computationally. In this paper, we demonstrate that inherent image phase smoothness of spin-echo MRI images could be utilized for separate filtering of real and imaginary complex image channels to achieve unbiased image denoising. The concept is demonstrated with a novel nonlinear diffusion filtering scheme developed for complex image filtering. In our proposed method, the separate diffusion processes are coupled through combined diffusion coefficients determined from the image magnitude. The new method has been validated with simulated and real MRI data. The new method has provided efficient denoising and bias removal in conventional and black-blood angiography MRI images obtained using fast spin echo acquisition protocols.

Medical imaging techniques often suffer from low signal-to-noise ratio (SNR)/signal-to-contrast ratio (CNR). Filtering is an effective technique commonly used to improve the SNR/CNR. A fundamental difficulty in filtering is its smearing effect on important structures along with the suppression of noise. Anisotropic diffusion attempts to partially arrest smoothing across boundaries, which are recognized by high gradients. Improvements to this approach were demonstrated by bringing local "object scale" into the framework that effects intense filtering inside homogeneous regions while constricting the process inside regions containing fine details. A major drawback of previously used spherical object scale concept is its failure to recognize structural orientation and anisotropy. Therefore, the method failed to smooth along an edge or a thin elongated structure. Recently, the notion of tensor scale (in short, t-scale)-a local morphometric descriptor-has been introduced that provides a unified description of local structure size, orientation, and anisotropy. At any point its t-scale is the parametric representation of the largest ellipse (ellipsoid in 3D) centered at that point and contained inside the same homogeneous region. In this paper, a new anisotropic diffusive filtering method is presented, wherein at any location, along any direction, the degree of filtering is governed by t-scale that restricts filtering across boundaries and fine structures while encouraging smoothing along edges or elongated structures. Results of applications of the proposed method on several phantoms and real medical images are presented and its effectiveness is evaluated by comparing it with both spherical scale-based and ordinary diffusive filtering.

Scale is a fundamental concept useful in almost all image processing and analysis tasks including segmentation, filtering, interpolation, registration, visualization, and quantitative analysis. Broadly speaking, scale related work can be divided into three categories: (1) multi-scale or scale-space representation, (2) local scale, and (3) locally adaptive scale. The original formulation of scale in the form of scale-space theory came from the presence of multiple scales in nature and the desire to represent measured signals at multiple scales. However, since this representation did not suggest how to select appropriate scales, the notion of local scale was proposed to pick the right scale for a particular application from the multi-scale representation of the image. Recently, there has been considerable interest in developing locally adaptive scales, the idea being to consider the local size of object in carrying out whatever local operations that are to be done on the image. However, existing locally adaptive models are limited by shape, size, and anisotropic constraints. In this work, we propose a generalized scale model which is adaptive like other local morphometric models, and yet possesses the global spirit of multi-scale representations. We postulate that this semi-locally adaptive nature of generalized scale confers it certain distinct advantages over other global and local scale formulations. Further, generalized scale can be easily applied to solve a range of image processing problems. One such problem that we address in this paper is inhomogeneity correction in MR images. We qualitatively and quantitatively demonstrate the superiority of our generalized scale-based correction method over an existing scale-based correction technique, while retaining all the advantages of the existing scale-based method over those published in the literature.

Breast cancer is second only to lung cancer as a tumor-related cause of death in women. Currently, the method of choice for the early detection of breast cancer is mammography. While sensitive to the detection of non palpable breast lesions, its positive predictive value (PPV) is low, resulting in biopsies that are only 15%-34% likely to reveal malignancy. This paper explores the use of a recently designed Support Vector Machine (SVM)/Generalized Regression Neural Network (GRNN) Oracle hybrid to classify breast lesions and evaluate the software's performance as an interpretive aid to radiologists. The main objective of the research was to perform an independent analysis, using a new, integrated film screen mammogram data base of approximately 2500 cases from five separate institutions, to verify results obtained previously[14]. This study demonstrated the following: (1) The DE crossover constant has little, if any, effect on measures of performance (MOP). (2) A specificity of approximately 5.6% is achieved at 100% sensitivity, which increases to approximately 36% at 95% sensitivity. (3) PPV increases from 51% to 56% as sensitivity is decreased from 100 to 95%, respectively.

Breast MR is being employed to detect, diagnose, and stage breast cancer. With a breast MR study, areas that exhibit rapid uptake of contrast followed by washout behavior have been shown to be indicative of malignant tissue. The most common way to display this temporal information is with a time versus percent enhancement curve that plots the enhancement of the tissue for each series relative to a baseline or pre contrast series. The generation of time curves is commonly done using manual methods, but could easily be automated by a computer to reduce user variability. The information obtained by the time curve can then be used for computer assisted analysis of suspicious areas. We compare two methods for the automated detection of such time curves for 42 malignant lesions. The first method is a previously published technique which finds the highest enhancing 3x3 area of a lesion to generate a curve. The second method is a new hierarchical search that examines end time point behavior in combination with enhancement to find an optimal curve location. The two methods for curve generation are examined in their ability to produce a washout type curve, which has a greater likelihood of being malignant than curves that continue to enhance. The time curves found using highest percent enhancement method showed washout in 52 percent (22/42) of lesions. Using the hierarchical search algorithm, 90 percent (38/42) of lesions showed washout type curves

The objective of this research is to evaluate and compare the performance of our automated detection algorithm on isolated and attached nodules in whole lung CT scans. Isolated nodules are surrounded by the lung parenchyma with no attachment to large solid structures such as the chest wall or mediastinum surface, while attached nodules are adjacent to these structures. The detection algorithm involves three major stages. First, the region of the image space where pulmonary nodules are to be found is identified. This involves segmenting the lung region and generating the pleural surface. In the second stage, which is the hypothesis generation stage, nodule candidate locations are identified and their sizes are estimated. The nodule candidates are successively refined in the third stage a sequence of filters of increasing complexity. The algorithm was tested on a dataset containing 250 low-dose whole lung CT scans with 2.5mm slice thickness. A scan is composed of images covering the whole lung region for a single person. The dataset was partitioned into 200 and 50 scans for training and testing the algorithm. Only solid nodules were considered in this study. Experienced chest radiologists identified a total of 447 solid nodules. 345 and 102 of the nodules were from the training and testing datasets respectively. 126(28.2%) of the nodules in the dataset were attached nodules. The detection performance was then evaluated separately for isolated and attached nodule types considering different size ranges. For nodules 3mm and larger, the algorithm achieved a sensitivity of 97.8% with 2.0 false positives (FPs) per scan and 95.7% with 19.3 FPs per scan for isolated and attached nodules respectively. For nodules 4mm and larger, a sensitivity of 96.6% with 1.5 FP per scan and a 100% sensitivity with 13 FPs per scan were obtained for isolated and attached nodule types respectively. The results show that our algorithm detects isolated and attached nodules with comparable sensitivity but differing number of false positives per scan. The high number of false positives for attached nodule detection was mainly due to the complexity of the mediastinum lung surface.

The purpose of this paper is to present a new segmentation routine developed for mammographic masses. We previously developed a computer-aided detection (CAD) system for mammographic masses that employed a simple but imprecise segmentation procedure. To improve the systems performance, an iterative, linear segmentation routine was developed. The routine begins by employing a linear discriminant function to determine the optimal threshold between estimates of an objects interior and exterior pixels. After applying the threshold and identifying the objects outline, two constraints are applied to minimize the influence of extraneous background structures. Each iteration further refines the outline until the stopping criterion is reached. The segmentation algorithm was tested on a database of 181 mammographic images that contained forty-nine malignant and fifty benign masses. A set of suspicious regions of interest (ROIs) was found using the previous CAD system. Twenty features were measured from the regions before and after applying the new segmentation routine. The difference in the features discriminatory ability was examined via receiver operating characteristic (ROC) analysis. A significant performance difference was observed in many features, particularly those describing the object border. Free-response ROC (FROC) curves were utilized to examine how the overall CAD system performance changed with the inclusion of the segmentation routine. The FROC performance appeared to be improved, especially for malignant masses. When detecting 90% of the malignant masses, the previous system achieved 4.4 false positives per image (FPpI) compared to the post-segmentation systems 3.7 FPpI. At 85%, the respective FPpI are 4.1 and 2.1.

The purpose of the study was to develop and evaluate a content-based image retrieval (CBIR) approach for computer-assisted diagnosis of masses detected in screening mammograms. The system follows an information theoretic retrieval scheme with a BIRADS-based relevance feedback (RF) algorithm. Initially, a knowledge databank of 365 mammographic regions of interest (ROIs) was created. They were all 512x512 pixel ROIs extracted from DDSM mammograms digitized using the Lumisys digitizer. The ROIs were extracted around the known locations of the annotated masses. Specifically, there were 177 ROIs depicting a biopsy-proven malignant mass and 188 ROIs with a benign mass. Subsequently, the CBIR algorithm was implemented using mutual information (MI) as the similarity metric for image retrieval. The CBIR algorithm formed the basis of a knowledge-based CAD system. Given a databank of mammographic masses with known pathology, a query mass was evaluated. Based on their information content, all similar masses in the databank were retrieved. A relevance feedback algorithm based on BIRADS findings was implemented to determine the relevance factor of the retrieved masses. Finally, a decision index was calculated using the query's k best matches. The decision index effectively combined the similarity metric of the retrieved cases and their relevance factor into a prediction regarding the malignancy status of the mass depicted in the query ROI. ROC analysis was to evaluate diagnostic performance. Performance improved dramatically with the incorporation of the relevance feedback algorithm. Overall, the CAD system achieved ROC area index AZ= 0.86±0.02 for the diagnosis of masses in screening mammograms.

In this work, we present a computer-aided detection (CAD) algorithm for small lung nodules on low-dose MSCT images. With this technique, identification of potential lung nodules is carried out with a local density maximum (LDM) algorithm, followed by reduction of false positives from the nodule candidates using task-specific 2-D/3-D features along with a knowledge-based nodule inclusion/exclusion strategy. Twenty-eight MSCT scans (40/80mAs, 120kVp, 5mm collimation/2.5mm reconstruction) from our lung cancer screening program that included at least one lung nodule were selected for this study. Two radiologists independently interpreted these cases. Subsequently, a consensus reading by both radiologists and CAD was generated to define a “gold standard”. In total, 165 nodules were considered as the “gold standard” (average: 5.9 nodules/case; range: 1-22 nodules/case). The two radiologists detected 146 nodules (88.5%) and CAD detected 100 nodules (60.6%) with 8.7 false-positives/case. CAD detected an additional 19 nodules (6 nodules > 3mm and 13 nodules < 3mm) that had been missed by both radiologists. Preliminary results show that the CAD is capable of detecting small lung nodules with acceptable number of false-positives on low-dose MSCT scans and it can detect nodules that are otherwise missed by radiologists, though a majority are small nodules (< 3mm).

We previously developed a scheme to automatically detect pulmonary nodules on CT images, as a part of computer-aided diagnosis (CAD) system. The proposed method consisted of two template-matching approaches based on simple models that simulate real nodules. One was a new template-matching technique based on a genetic algorithm (GA) template matching (GATM) for detecting nodules within the lung area. The other one was a conventional template matching along the lung wall [lung wall template matching (LWTM)] for detecting nodules on the lung wall. After the two template matchings, thirteen feature values were calculated and used for eliminating false positives. Twenty clinical cases involving a total of 557 sectional images were applied; 71 nodules out of 98 were correctly detected with the number of false positives at approximately 30.8/case by applying two template matchings (GATM and LWTM) and elimination process of false positives. In this study, five features were newly added, and threshold-values of our previous features were reconsidered for further eliminating false positives. As the result, the number of false positives was decreased to 5.5/case without elimination of true positives.

We developed a two-stage computerized mass detection algorithm for digital tomosynthesis images of the breast. Rather than analyze the reconstructed 3D breast volume, our algorithm operates on each of the 2D projection images directly. We chose this approach because reconstruction algorithms for breast tomosynthesis are still being optimized, which can alter the appearance of the 3D reconstructed breast volume. Furthermore this approach allows us to take advantage of mass detection methods already developed for conventional two-view projection mammography, which are similar to projection images for digital tomosynthesis. We applied our algorithm to two tomosynthesis image sets, one of which was a computer simulated 3D breast phantom, and one was a clinical image set. In both cases, the lesion was detected in the first stage of the algorithm, while the second stage of the algorithm efficiently reduced false positive detections.

We developed a new method for automated detection of colonic polyps in CT colonography. The colon is extracted from CT images by use of a centerline-based colon segmentation method. Polyp candidates are detected by use of hysteresis thresholding and fuzzy merging. The regions of the polyp candidates are segmented by use of conditional morphological dilation. False-positive polyp candidates are reduced by a region-based supine-prone correspondence method and by a Bayesian neural network with shape and texture features. To evaluate the method, CT colonography was performed for 121 patients with standard technique and single- and multi-detector helical scanners by use of 2.5-5.0 mm collimations, 1.0-5.0 mm reconstruction intervals, and 60-100 mA tube currents. Twenty-eight patients had a total of 42 polyps: 22 polyps were 5-10 mm, and 20 polyps were 11-25 mm in size. A leave-one-out evaluation of the CAD scheme with by-patient elimination yielded 93% by-polyp and by-patient detection sensitivities with 2.0 false-positive detections per data set on average. The average computation time was 4 minutes per data set. The results indicate that the CAD scheme may be useful in improving the performance of computer-aided detection for colon cancer in a clinical screening setting.

We developed a computer-aided diagnosis (CAD) scheme to detect and quantitatively assess interstitial lung diseases (ILD) depicted on low-dose and multi-slice helical high-resolution computed tomography (CT) examinations. Eighteen CT cases acquired from patients who underwent routine low-dose whole-lung screening examinations for the detection of lung cancer were used to test the scheme. ILD was identified in all of these cases. The CAD scheme involves multiple steps to segment lung areas, identify suspicious ILD regions depicted on each CT slice, and generate volumetric ILD lesions by grouping and matching ILD regions detected on multiple adjacent slices. The scheme computes five “global” features for each identified ILD region, which include size (or volume), contrast, average local pixel value fluctuation, mean of stochastic fractal dimension, and geometric fractal dimension. Two sets of classification rules are applied to remove false-positive detections. The severity of ILD in each case was rated by one experienced chest radiologist into one of the three categories (mild, moderate, and severe). A distance-weighted k-nearest neighbor algorithm and round-robin validation method was applied to classify each testing case into one of the three categories of severity. In this experiment, the CAD scheme classified 78% (14 out of 18) cases into the same categories as rated by the radiologist.

A decision support system has been developed to assist the radiologist during mammogram classification. In this paper, mass identification and segmentation methods are discussed in brief. Fuzzy region-growing techniques are applied to effectively segment the tumour candidate from surrounding breast tissue. Boundary extraction is implemented using a unit vector rotating about the mass core. The focus of this work is on the feature extraction and classification processes. Important information relating to the malignancy of a mass may be derived from its morphological properties. Mass shape and boundary roughness are primary features used in this research to discriminate between the two types of lesions. A subset from thirteen shape descriptors is input to a binary decision tree classifier that provides a final diagnosis of tumour malignancy. Features that combine to produce the most accurate result in distinguishing between malignant and benign lesions include: spiculation index, zero crossings, boundary roughness index and area-to-perimeter ratio. Using this method, a classification result of high sensitivity and specificity is achieved, with false-positive and falsenegative rates of 9.3% and 0% respectively.

In this paper we apply multiscale entropy (MSE) analysis to data obtained from magnetic resonance imaging of the female breast. All cases include lesions that were histologically proven as malignant tumors. Our results indicate that multiscale entropy analysis can play an important role in the detection of tumor tissue when applied to single datasets, but does not allow to calculate universal morphological features. The performance of MSE was examined with respect to traditional features such as difference imaging.

Mammography is an effective tool for the early detection of breast cancer; however, most women referred for biopsy based on mammographic findings do not, in fact, have cancer. This study is part of an ongoing effort to reduce the number of benign cases referred for biopsy by developing tools to aid physicians in classifying suspicious lesions. Specifically, this study examines the use of an Evolutionary Programming (EP)-derived Support Vector Machine (SVM) with a modified radial basis function (RBF) kernel, and compares this with results using a normal Gaussian radial basis function kernel. Results demonstrate that the modified kernel can provide moderate performance improvements; however, due to its ability to create a more complex decision surface, this kernel can easily begin to memorize the training data resulting in a loss of generalization ability. Nonetheless, these methods could reduce the number of benign cases referred for biopsy by over half, while missing less than 5% of malignancies. Future work will focus on methods to improve the EP process to preserve SVMs which generalize well.

Dynamic contrast breast MRI is becoming an important adjunct in screening women at high risk for breast cancer, determining extent of disease (staging) and monitoring response to therapy. In dynamic contrast breast MRI, regions of rapid contrast uptake indicate increases in vascularity which can be associated with abnormal tissue, sometimes significant for malignant disease. To show these areas of enhancement, subtractions between the pre and post contrast images and maximum intensity projections (MIPs) are computed. Many projections are obscured by normally enhancing anatomy (heart, aorta, pulmonary vessels). Identification of these structures allows their removal from MIPs, which improves image quality, diagnostic utility and the conspicuity of the enhancing regions. In this study, a fully automated classifier is presented which uses the spatial location of enhancing regions to separate those that occur inside the chest wall from those occurring in the tissue of interest (breast, axilla, chest wall). The classifier was trained on 21 studies each acquired at a different institution (699 clusters of pixels), and tested on 7 studies (231 clusters of pixels) that were not part of the training set. Multiple cost functions for training were examined. The measurements for the peak performance of the classifier were sensitivity 97.0%, specificity 99.4%, PPV 99.9%, NPV 78.8%.

Recently, the multi-slice CT for a lung cancer screening has 10 times as high resolution as the conventional helical CT. Small nodules can be found in this CT image. The diagnostic ability of expert physicians will improve by using this CT images. We get the CT image of 2mm thickness and 10 mm thickness from each subject. In this paper, we evaluated about the reading test for the lung cancer by expert physicians, using two kinds of images.

A computerized scheme for early severe acute respiratory syndrome(SARS) lesion detection in digital chest radiographs is presented in this paper. The total scheme consists of two main parts: the first part is to determine suspect lesions by the theory of locally orderless images(LOI) and their spatial features; the second part is to select real lesions among these suspect ones by their frequent features. The method we used in the second part is firstly developed by Katsuragawa et al with necessary modification. Preliminary results indicate that these features are good criterions to tell early SARS lesions apart from other normal lung structures.

An automatic extraction of pulmonary emphysema area on 3-D chest CT images was performed using an adaptive thresholding technique. We proposed a method to estimate the ratio of the emphysema area to the whole lung volume. We employed 32 cases (15 normal and 17 abnormal) which had been already diagnosed by radiologists prior to the study. The ratio in all the normal cases was less than 0.02, and in abnormal cases, it ranged from 0.01 to 0.26. The effectiveness of our approach was confirmed through the results of the present study.

As well as mass and microcalcification, architectural distortion is a very important finding for the early detection of breast cancer via mammograms, and such distortions can be classified into three typical types: spiculation, retraction, and distortion. The purpose of this work is to develop an automatic method for detecting areas of architectural distortion with spiculation. The suspect areas are detected by concentration indexes of line-structures extracted by using mean curvature. After that, discrimination analysis of nine features is employed for the classifications of true and false positives. The employed features are the size, the mean pixel value, the mean concentration index, the mean isotropic index, the contrast, and four other features based on the power spectrum. As a result of this work, the accuracy of the classification was 76% and the sensitivity was 80% with 0.9 false positives per image in our database in regard to spiculation. It was concluded that our method was effective in detectiong the area of architectural distortion; however, some architectural distortions were not detected accurately because of the size, the density, or the different appearance of the distorted areas.

We have developed an algorithm that can be used to distinguish the central part of the vertebral body from an abdominal X-ray CT image and to automatically calculate three measures to diagnose the degree of osteoporosis in a patient. In addition, we examined whether it is possible to use these CT images as an aid in diagnosing osteoporosis. Three measures that were automatically extracted from the central part of a vertebral body in the CT images were compared with the bone mineral density (BMD) values that were obtained from the same vertebral body. We calculated the mean CT number, coefficient of variation, and the first moment of power spectrum in the recognized vertebral body. We judged whether a patient had osteoporosis using the diagnostic criteria for primary osteoporosis (Year 2000 revision, published by the Japanese Society for Bone and Mineral Research). We classified three measures for normal and abnormal groups using the principal component analysis, and the two groups were compared with the results obtained from the diagnostic criteria. As a result, it was found that the algorithm could be used to distinguish the central part of the vertebral body in the CT images and to calculate these measures automatically. When distinguishing whether a patient was osteoporotic or not with the three measures obtained from the CT images, the ratio (sensitivity) usable for diagnosing a patient as osteoporotic was 0.93 (14/15), and the ratio (specificity) usable for diagnosing a patient as normal was 0.64 (7/11). Based on these results, we believe that it is possible to utilize the measures obtained from these CT images to aid in diagnosing osteoporosis.

A technique is presented for segmentation and quantification of stenosed internal carotid arteries in three-dimensional contrast-enhanced magnetic resonance angiography. Segmentation with sub-voxel accuracy of the internal carotid arteries (ICAs) has been achieved via level-set techniques in which the central axis serves as initialization. The central axis is determined with minmal user-interaction, viz. two user-defined points. Quantification is performed by measuring the cross-sectional area in the stenosis and at a reference segment in planes perpendicular to the central axis. The technique was applied to 52 ICAs. It is demonstrated that the method's reproducibility is better than the intra-observer agreement. Furthermore, the agreement between the presented method and the observers is better than the inter-observer agreement.

The features used in the ABCD rule for characterization of skin lesions suggest that the spatial and frequency information in the nevi changes at various stages of melanoma development. To analyze these changes wavelet transform based features have been reported. The classification of melanoma using these features has produced varying results. In this work, all the reported wavelet transform based features are combined to form a single feature set. This feature set is then optimized by removing redundancies using principal component analysis. A feed forward neural network trained with the back propagation algorithm is then used in the classification process to obtain better classification results.

The purpose of this study is to develop a computerized scheme for the discrimination between benign and malignant clustered microcalcifications that would aid radiologists in interpreting mammograms. In our scheme, microcalcifications in regions of interest (ROIs) are detected by using morphological filter. Then, four feature values including the total number, mean area, mean circularity and mean minimum distance of microcalcifications are calculated for classification. Gaussian-distributed membership functions used for fuzzy logic are determined from means and standard deviations of these feature values. Finally, fuzzy logic using the genetic-algorithm for optimization of membership functions is employed to classify clustered microcalcifications in unknown ROI. Our scheme was applied to twenty mammographic images with microcalcifications in the Mammographic Image Analysis Society database, containing thirteen benign and twelve malignant ROIs. Of the images ten each benign and malignant ROIs were used for training in fuzzy logic. The remaining five images were classified as benign or malignant cases by fuzzy logic. All sets of their combinations were employed to obtain the result. As the results, the average accuracy was approximately 88% (sensitivity: 100%, specificity: 77%), and Az value of ROC curve was 0.95.