Paulo S. Rodrigues

Non-Extensive Entropy for CAD Systems of Breast Cancer Images

Recent statistics show that breast cancer is a major cause of death among women in all of the world. Hence, early diagnostic with computer aided diagnosis (CAD) systems is a very important tool. This task is not easy due to poor ultrasound resolution and large amount of patient data size. Then, initial image segmentation is one of the most important and challenging task. Among several methods for medical image segmentation, the use of entropy for maximization the information between the foreground and background is a well known and applied technique. But, the traditional Shannon entropy fails to describe some physical systems with characteristics such as long-range and longtime interactions. Then, a new kind of entropy, called non-extensive entropy, has been proposed in the literature for generalizing the Shannon entropy. In this paper, we propose the use of non-extensive entropy, also called q-entropy, applied in a CAD system for breast cancer classification in ultrasound of mammographic exams. Our proposal combines the non-extensive entropy, a level set formulation and a support vector machine framework to achieve better performance than the current literature offers. In order to validate our proposal, we have tested our automatic protocol in a data base of 250 breast ultrasound images (100 benign and 150 malignant). With a cross-validation protocol, we demonstrate systems accuracy, sensitivity, specificity, positive predictive value and negative predictive value as: 95%, 97%, 94%, 92% and 98%, respectively, in terms of ROC (receiver operating characteristic) curves and Az areas

Automated carotid artery intima layer regional segmentation

Evaluation of the carotid artery wall is essential for the assessment of a patients cardiovascular risk or for the diagnosis of cardiovascular pathologies. This paper presents a new, completely user-independent algorithm called carotid artery intima layer regional segmentation (CAILRS, a class of AtheroEdge™ systems), which automatically segments the intima layer of the far wall of the carotid ultrasound artery based on mean shift classification applied to the far wall. Further, the system extracts the lumen-intima and media-adventitia borders in the far wall of the carotid artery. Our new system is characterized and validated by comparing CAILRS borders with the manual tracings carried out by experts. The new technique is also benchmarked with a semi-automatic technique based on a first-order absolute moment edge operator (FOAM) and compared to our previous edge-based automated methods such as CALEX (Molinari et al 2010 J. Ultrasound Med. 29 399-418, 2010 IEEE Trans. Ultrason. Ferroelectr. Freq. Control 57 1112-24), CULEX (Delsanto et al 2007 IEEE Trans. Instrum. Meas. 56 1265-74, Molinari et al 2010 IEEE Trans. Ultrason. Ferroelectr. Freq. Control 57 1112-24), CALSFOAM (Molinari et al Int. Angiol. (at press)), and CAUDLES-EF (Molinari et al J. Digit. Imaging (at press)). Our multi-institutional database consisted of 300 longitudinal B-mode carotid images. In comparison to semi-automated FOAM, CAILRS showed the IMT bias of -0.035 ± 0.186 mm while FOAM showed -0.016 ± 0.258 mm. Our IMT was slightly underestimated with respect to the ground truth IMT, but showed uniform behavior over the entire database. CAILRS outperformed all the four previous automated methods. The systems figure of merit was 95.6%, which was lower than that of the semi-automated method (98%), but higher than that of the other automated techniques.

A new methodology based on q-entropy for breast lesion classification in 3-D ultrasound images.

Classification of breast lesions is clinically most relevant for breast radiologists and pathologists for early breast cancer detection. This task is not easy due to poor ultrasound resolution and large amount of patient data size. This paper proposes a five step novel and automatic methodology for breast lesion classification in 3-D ultrasound images. The first three steps yield an accurate segmentation of the breast lesions based on the combination of (a) novel non-extensive entropy, (b) morphologic cleaning and (c) accurate region and boundary extraction in level set framework. Segmented lesions then undergo five feature extractions consisting of: area, circularity, protuberance, homogeneity, and acoustic shadow. These breast lesion features are then input to a support vector machine (SVM)-based classifier that classifies the breast lesions between malignant and benign types. SVM utilizes B-spline as a kernel in its framework. Using a data base of 250 breast ultrasound images (100 benign and 150 malignant) and utilizing the cross-validation protocol, we demonstrate systems accuracy, sensitivity, specificity, positive predictive value and negative predictive value as: 95%, 97%, 94%, 92% and 98% respectively in terms of ROC curves and Az areas, better in performance than the current literature offers

Computing the q-index for Tsallis Nonextensive Image Segmentation

The concept of entropy based on Shannon Theory of Information has been applied in the field of image processing and analysis since the work of T. Pun [1]. This concept is based on the traditional Boltzaman-Gibbs entropy, proposed under the classical thermodynamic. On the other hand, it is well known that this old formalism fails to explain some physical system if they have complex behavior such as long rang interactions and long time memories. Recently, studies in mechanical statistics have proposed a new kind of entropy, called Tsallis entropy (or non-extensive entropy), which has been considered with promising results on several applications in order to explain such phenomena. The main feature of Tsallis entropy is the

Distributed Visualization Using VTK in Grid Environments

q

Statistical learning approaches for discriminant features selection

-index parameter, which is close related to the degree of system nonextensivity. In 2004 was proposed[2] the first algorithm for image segmentation based on Tsallis entropy. However, the computation of the q-index was already an open problem. On the other hand, in the field of image segmentation it is not an easy task to compare the quality of segmentation results. This is mainly due to the lack of an image ground truth based on human reasoning. In this paper, we propose the first methodology in the field of image segmentation for q-index computation and compare it with other similar approaches using a human based segmentation ground truth. The results suggest that our approach is a forward step for image segmentation algorithms based on Information Theory.

Analyzing natural human language from the point of view of dynamic of a complex network

In this paper we focus on distributed visualization using the visualization toolkit (VTK) in grid environments. We propose a distributed architecture, based on data parallelism, that allows the distribution of visualization tasks over a grid environment. We decided for globus toolkit as a middleware to provide access and location transparencies. We also add facilities for dynamic allocation of resources by using a Java framework. The focused visualization technique is Laplacian smoothing which is provided by a specific filter of the VTK library. We emphasize the obtained speedup in the experiments and discuss the implementation of pipeline parallelism as well as the generalization of our architecture for other VTK applications.

CARES 3.0: A two stage system combining feature-based recognition and edge-based segmentation for CIMT measurement on a multi-institutional ultrasound database of 300 images

Supervised statistical learning covers important models like Support Vector Machines (SVM) and Linear Discriminant Analysis (LDA). In this paper we describe the idea of using the discriminant weights given by SVM and LDA separating hyperplanes to select the most discriminant features to separate sample groups. Our method, called here as Discriminant Feature Analysis (DFA), is not restricted to any particular probability density function and the number of meaningful discriminant features is not limited to the number of groups. To evaluate the discriminant features selected, two case studies have been investigated using face images and breast lesion data sets. In both case studies, our experimental results show that the DFA approach provides an intuitive interpretation of the differences between the groups, highlighting and reconstructing the most important statistical changes between the sample groups analyzed.

Using Tsallis entropy into a Bayesian network for CBIR

We present complex network model for semantic representation of human language.Contextual scope from texts are inferred from physical measures in complex networks.The complex network modeled presents behavior of scale-free networks.Human language has modular characteristic in the relationship between words.The relationship between words of human language is a sparse space. With increasing amount of information, mainly due to the explosive growth of Internet, the demand for applications of automatic text analysis has also grown. One of the tools that has increased in importance in the understanding of problems related to this area are complex networks. This tool merges graph theory and statistical methods for modeling important problems. In several research fields, complex networks are studied from the various points of view, such as: topology of networks, extraction of physical features and statistics, specific applications, comparison of metrics and study of physical phenomena. Linguistic is one area that has received great attention, particularly due to its close relationship with issues arising from the emergence of large text databases. Thus, many studies have emerged for modeling of complex networks in this area, increasing the demand for efficient algorithms for feature extraction, network dynamic observation and comparison of behavior for different types of languages. Some works for specific languages such as English, Chinese, French, Spanish, Russian and Arabic, have discussed the semantic aspects of these languages. On the other hand, as an important feature of a network we can highlight the computation of average clustering coefficient. This measure has a physical impact on the network topology studies and consequently on the conclusions about the semantics of a language. However its computational time is of O(n3), making its computing prohibitive for large current databases. This paper presents as main contribution a modeling of two complex networks: the first one, in English, is constructed from a specific medical database; the second, in Portuguese, from a journalistic manually annotated database. Our paper then presents the study of the dynamics of these two networks. We show their small-world behavior and the influence of hubs, suggesting that these databases have a high degree of Modularity, indicating specific contexts of words. Also, a method for efficient clustering coefficient computation is presented, and can be applied to large current databases. Other features such as fraction of reciprocal connections and average connection density are also calculated and discussed for both networks.

Describing patterns in flow-like images

The intima-media thickness of the carotid artery (CIMT) is a validated marker of atherosclerosis. Accurate CIMT measurement can be performed by specifically designed computer algorithms. We improved a previous CIMT measurement technique by introducing a smart heuristic search for the lumen-intima (LI) and media-adventitia (MA) interfaces of the carotid distal wall. We called this new release as CARES 3.0 (a class of AtheroEdge™ system, a patented technology from Global Biomedical Technologies, Inc., CA, USA). CARES 3.0 is completely automated and adopts an integrated approach for carotid location in the image frame, followed by segmentation based on edge snapper and heuristic search. CARES 3.0 was benchmarked against three other techniques on a 300 image multi-institutional database. One of the techniques was user-driven. The CARES 3.0 CIMT measurement bias was −0.021±0.182 mm, which was better than that of the semi automated method (−0.036±0.183 mm). CARES 3.0 outperformed the other two fully automated methods. The Figure-of-Merit of CARES 3.0 was 97.4%, better than that of the semi-automated technique (95.4%).