Exploring Deep-Based Approaches for Semantic Segmentation of Mammographic Images

Abstract

Pectoral muscle and background elimination are common steps for automated software in mammographic image preprocessing. We investigate FCNs, U-nets and SegNets in the task of mammogram segmentation, addressing three subtasks: pectoral muscle, background and breast region segmentation. The MIAS and INbreast datasets were used for evaluating Deep Neural Networks on the segmentation of these regions. Several objective evaluation metrics were used in order to compare our results with the ones available in the literature. State-of-the-art results were observed in most comparisons, significantly surpassing the baselines in most metrics. Best Jaccard values (in %) for Deep Learning algorithms were \(89.7\pm 2.5\), \(98.4\pm 0.1\) and \(97.0\pm 0.4\) for pectoral muscle, background and breast region segmentation, respectively, in the MIAS dataset. For INbreast, the best Jaccard value achieved for pectoral muscle segmentation was \(90.8\pm 2.5\).