Computer Aided Diagnosis for Breast Lesions in Ultrasound Images Under Non-Extensive Tsallis Entropy

This paper presents a methodology to study breast lesions captured from ultrasound (US) devices. One of the main problems behind this theme is the separation of the lesion from their background in order to further extract their features, allowing a better automatic analysis later. Previous works have shown that by considering such images as non-extensive physical systems allows implementation of specific algorithms which facilitates the lesion separation. Then, as Tsallis entropy has appeared as a promise theory underlining non-extensive systems, in this paper we propose a recursive algorithm, based on this theory, in order to apply optimal threshold for breast lesion ultrasound image segmentation. A fundamental issue about segmentation methods that use Tsallis statistics is the setting of the parameter q, also called non-extensivity parameter. Recently, we introduced a method for automatic computation of q value in general natural images. Now, our strategy is to use this automatic computation combined with two levels of Tsallis segmentation, specific for breast lesion ultrasound images. Then, experimentally, we show that our methodology properly extracts the lesion from the background allowing application of specific heuristics which delimit the lesion core for further analysis of the following three different lesion features: circularity, contrast and acoustic shadow/reinforcement. We meet the medical literature when we show that malignant lesions have high acoustic reinforcement behavior and benign ones have smoothing circularity as well. Our experiments have achieved optimal values for a data base of 250 images, and objective index rates that are better than those reported in the current literature under the same conditions. Two of these indices are: Sensitiveness = 1.0 and Specificity = 1.0, which are reached when we use acoustic shadow/reinforcement as the main lesion’s feature. The malignant/benign classification is accomplished through a SVM classifier with a B-Spline kernel.