Segmentation of Mammogram Images Using Deep Learning for Breast Cancer Detection

Abstract

Breast Cancer is one of the most common forms of cancer occurring in women. According to GLOBCON, 24.2% of new cancer types are related to breast cancer. Early detection followed by a proper diagnosis can help to combat this kind of serious disease. Segmentation of mammography lesions has been proven to be a valuable source of information for the detection and classification of breast cancer. Segmentation can assist both in extracting shape-related features and providing accurate localization of the lesions. In this paper, detailed research on the segmentation of the mammography images using two different networks, namely U-Net and BCDU-Net is done. A comparison of the performance concerning Dice Coefficient and Jaccard Similarity is done on two subsets of the same dataset, one on the full mammogram image and another on the ROI extracted from the mammogram. The evaluation results of the proposed research achieve the highest Jaccard and Dice coefficient of 0.7872 and 0.8376 respectively on the full mammogram. Whereas the figures for ROI segmentation are 0.8098 and 0.8723 respectively. These results demonstrate that, with equal data size, both the network provides better segmentation performance on ROI than on full mammogram.