Diagnostic Assessment of Deep Learning Algorithms for Detection and Segmentation of Lesion in Mammographic Images

Computer-aided detection or diagnosing support methods aims to improve breast cancer screening programs by helping radiologists to evaluate digital mammography (DM) exams. This system relates to the use of deep learning for automated detection and segmentation of soft tissue lesions at the early stage. This paper presents a novel deep learning approach, based on a two stage object detector combining an enhanced Faster R-CNN with the Libra R-CNN structure for the Object Detection segment. A segmentation network is placed on top of previous structure in order to provide accurate extraction and localization of masses various features, i.e: margin, shape. The segmentation head is based on a Recurrent Residual Convolutional Neural Network and can lead to an additional feature classification for specific instance properties. A database of digital mammograms was collected from one vendor, Hologic, of which 1,200 images contained masses. The performance for our automated detection system was assessed with the sensitivity of the model which reached a micro average recall: 0.892, micro average precision: 0.734, micro average F1 score: 0.805. Macro average recall: 0.896, macro average precision: 0.819, macro average F1 score: 0.843. The segmentation performance for the same test set was evaluated to a mean IOU of 0.859.