|
Because of the unclear boundaries and different shapes and sizes of breast masses, the accuracy of using traditional computer-aided diagnosis systems is low and it is difficult to meet the clinical requirements of physicians. In this paper, we propose a breast mass detection algorithm based on the combination of YOLOv5 and improved coordinate attention, to meet the clinical requirements of high accuracy and real-time. First, a novel backbone feature extraction network is constructed by combining the underlying backbone network and attention mechanism to fully learn useful features and suppress irrelevant features, thus enhancing the feature expression capability. Then a multi-path aggregation network is designed as the neck of feature fusion to fully fuse the feature information at different levels. Validation experiments are conducted on the DDSM breast mass dataset, and the results show that the network can accurately detect masses of different scales in different backgrounds with better real-time performance. Compared with the base YOLOv5, the network improves by 2.3% in accuracy.
|
