Feature extraction for classification of breast tumor images using artificial organisms

This paper describes a method for classification of hematoxylin and eosin ( HE)-stained breast tumor images into benign or malignant using the adaptive searching ability of artificial organisms. Each artificial organism has so me attributes, such as, age, internal energy and coordinates. In addition, the artificial organism has a differentiation function for evaluating "mali gnant" or "benign" tumors and the adaptive behaviors of each artificial org anism are evaluated using five kinds of texture features. The texture featu re of nuclei regions in normal mammary glands and that of carcinoma regions in malignant tumors are treated as "self" and "non-self," respectively. Th is model consists of two stages of operations for detecting tumor regions, the learning and searching stages. At the learning stage, the nuclei region s are roughly detected and classified into benign or malignant tumors. At t he searching stage, the similarity of each organism's environment is invest igated before and after the movement for detecting breast tumor regions pre cisely. The method developed was applied to 21 cases of test images and the distinction between malignant and benign tumors by the artificial organism s was successful in all cases. The proposed method has the following advant ages: the texture feature values for the evaluation of tumor regions at the searching stage are decided automatically during the learning stage in eve ry input image. Evaluation of the environment, whether the target pixel is a malignant tumor or not, is performed based on the angular difference in e ach texture feature. Therefore, this model can successfully detect tumor re gions and classify the type of tumors correctly without affecting a wide va riety of breast tumor images, which depends on the tissue condition and the degree of malignancy in each breast tumor case.