Investigating automatic techniques in segmentation accuracy of masses in digital mammography images

Abstract

Many procedures have been developed to aid in the early detection and diagnosis of breast cancer. In this context, Computer-Aided Diagnosis (CAD) schemes were designed to provide to the specialist a reliable second opinion. In such schemes there is a complex step which corresponds to the segmentation since good structures classification is dependent on the features extracted from the segmented images. In this work we propose the use of several methods of automatic segmentation of breast lesions, such as: watershed, fuzzy c-means, k-means, Self-Organizing Map (SOM), Enhanced Independent Component Analysis Mixture Model (EICAMM) and level set. In order to evaluate which of them could provide more accurate results in segmenting breast masses segmented images were compared with those manually delimited by an experienced radiologist. Ten quantitative measures were obtained from the images. These segmentation techniques were applied on different types of lesions, including images corresponding to dense breasts. From the evaluation the level set technique has proved being more effective for the images set used to testing all the methods. It has registered a higher overlap rates in relation to the image segmented by the specialist as well as low rates of under and oversegmentation, reflecting in the high accuracy and low false-positive and error rates.