Mass Detection in Mammograms Using a Robust Deep Learning Model

Computer-aided detection (CADe) methods can help radiologists detect breast cancer in its early stage. However, several breast mass detection methods have been proposed, they still produce a noticeable number of false positives (i.e., normal breast tissue is wrongly detected as a mass) due to the variability in breast density and mass size. In this paper, an automated detection method is proposed to detect breast masses from mammographic images using a modified Faster R-CNN detector based on Inception-ResNet-v2 feature extractor with a squeeze and excitation block. The squeeze and excitation mechanism provides channel inter-dependencies inside the Inception-ResNet-v2 feature extractor, which helps to extract low contrast texture features in mammogram images. Both qualitative and quantitative comparisons using INbreast dataset show that the proposed method outperforms the state-of-the-art methods. The proposed method yields the highest true positive rate (98.5%) and the lowest detection time (4 seconds). A visualization of the mass detection results can be found at https://youtu.be/PP2OldECuPY.