Metaheuristics for Specialization of a Segmentation Algorithm for Ultrasound Images
提出一种无种子分割系统，通过模糊规则系统和两种生物启发算法（遗传算法和蚁群算法）最小化分割误差，适用于乳腺癌筛查、卵巢评估和麻醉超声等临床任务。
An unseeded segmentation system applied to ultrasound (US) imaging is presented, based on a compact segmentation algorithm. Its basic behavior is adapted by a region selection algorithm controlled by a region classification function, which scores the relevance of regions generated from the previous segmentation step. This approach results in a completely unseeded system. Its behavior, represented by a fuzzy rule system, is specialized for the present clinical applications by means of two different bioinspired approaches that minimize the segmentation error against a human expert asked to fulfill the same task. The first one is based on a real-valued genetic algorithm, whereas the second is based on an ant colony stigmergic algorithm. The two methodologies are tested and benchmarked on four data sets: 1) breast US images for carcinoma screening; 2) obs/gyn US for ovarian follicles assessment; and 3) two applications in anesthesiology US during brachial anesthesia. Results show that the proposed bioinspired approaches are well suited for these complex tasks and can be used as a straightforward methodology to adapt an image segmentation algorithm to fulfill recognition tasks. The system could be adapted to any application in US imaging that requires identification of anatomical districts, morphological structures, and any other region of interest related to the clinical practice.