DeepLIMa: Deep Learning Based Lesion Identification in Mammograms

Abstract

Mammography is a major technique for early detection of breast cancer, typically through detection of masses or calcifications. However, how to help radiologists efficiently recognize these lesions remains a challenging problem. In this paper, we propose comprehensive deep learning based solutions to respectively detect masses and segment calcifications in mammograms. To achieve the optimal mass detection performance, our method combines Faster R-CNN with Feature Pyramid Networks, Focal Loss, and Non-Local Neural Networks. We thoroughly compare the proposed method and competing methods on three public datasets and an in-house dataset. The best detection results on our in-house dataset are an average precision of 0.933 and a recall of 0.976. Regarding calcification segmentation, we design a series of pre-processing methods including window adjustment, breast region extraction and artifact removal to normalize mammograms. A U-Net model with group normalization is then applied to segment calcifications. The proposed method is validated on our in-house dataset using a newly designed evaluation metric. The experimental results have demonstrated the great potential for this task.