DM-YOLOv5 for ABUS Detection

Automatic breast ultrasound (ABUS) technology provides help for the diagnosis of breast cancer. Developing an accurate computer-aided diagnosis system is important to improve the efficiency and accuracy of diagnosis. In this paper, based on YOLOv5, we develop an ABUS tumor detection network assisted by dense FPN and weakly supervised segmentation. Specifically, the feature fusion strategy in YOLOv5 is replaced by a densely connected feature pyramid network, which can enhance the feature representation. To suppress the false positive results, the channel attention module is introduced. Besides, to further improve the detection performance, a segment branch is added, forming a multi-task learning network. The proposed method achieves 89.8% Recall and 0.163 FPs/S. Experimental results show that the proposed method outperforms the vanilla YOLOv5 model and other detectors.