Automatic segmentation of multiple lesions in ultrasound breast image

Abstract

Breast cancer is one of the leading cancer that affects woman all around the world. Nowadays ultra sound imaging technique is used to diagnose various cancer because of its non-ionizing, on-invasive, and cheap cost. Breast lesion region in ultrasound images are classified depending upon the contour, shape, size and textural features of the segmented region. Seed point is the initial step in segmentation of lesion regions and if that point is located outside the lesion region, it leads to wrong segmentation which results in misclassification of the lesion regions. To avoid this, most of the time the seed point is located manually. In order to avoid this manual intervention, we are proposing a novel method in locating the seed point and also segmenting the breast lesion region automatically. In this method, the image is processed with tan function for effective distinguishing of breast lesion and normal region. Then using the trained neural network, the seed point is automatically located inside the lesion region and from the seed point the region of the lesion is grown and segmented automatically. Most of the past works on automatic segmentation of lesion had concentrated only in single lesion region, but using this proposed method, we were able to automatically segment multiple lesion regions in the image. Outcome of the proposed method is to detect automatically and dynamically separate the lesion region in the range between 90% to 97.5% of images.