U-net based approach for pectoralis muscle segmentation in digital mammography

Accurate segmentation of the breast is a fundamental step in computer-aided diagnosis (CAD) systems for mammography. In particular, several tasks, such as the classification of breast density, evaluation of correct positioning of the breast, and automatic detection and classification of suspicious lesions, preliminarily require an accurate segmentation of the pectoralis muscle. This study aims to propose an automatic breast segmentation algorithm that combines traditional methods with Deep Learning methods limited only to the border region between the muscle and the breast. This type of approach allows for reducing the risk of having good overall accuracy in multi-class classification that does not reflect adequate accuracy with respect to small classes, such as the pectoralis muscle in a mammographic image. The U-Net network was therefore implemented on patches extracted along the straight line with which the muscle-breast edge was first estimated. The predicted patches are repositioned to perform an edge refinement and obtain the total breast mask, using histogram-based thresholding to segment the background from the breast. The results show Dice values equal to 0.848 ± 0.196 and Jaccard index equal to 0.774 ± 0.227 for the single patches, and Dice values equal to 0.971 ± 0.011 and Jaccard index equal to 0.944 ± 0.022 for the entire breast segmentation. • A semi-automatic approach for accurate breast segmentation in digital mammography. • Applying deep learning to the muscle’s boundary provides reliable performances. • Use of small input patches and binary segmentation improves computational efficiency. • Boundary-wise segmentation allows detecting performance drops due to class imbalance.