Detection of Microcalcifications in Mammograms Based on Hyper Faster R-CNN

Microcalcifications are one of the important indicators of breast cancer. Accurate breast microcalcification detection can effectively assist radiologists in early diagnosis. The Hyper Faster R-CNN network is a deep neural network specially proposed for the detection of microcalcifications in mammograms. The network uses ResNeSt as the feature extraction network that can extract richer semantic features from the image, and uses the FPN feature fusion mechanism to fuse high-resolution features at low levels and high-semantic features at high levels, thereby improving the network's detection accuracy of microcalcifications. Focal Loss is adopted as the loss function to alleviate the problem of imbalanced microcalcification samples. To evaluate the defection effect of the proposed network on microcalcifications, the DDSM database is used to generate the microcalcification image dataset required for experiments, and the obtained AP value is up to 90.75%. In the end, the HF R-CNN network is compared with other state-of-the-art object detection networks. The experimental results show that the proposed HF R-CNN network can detect microcalcifications in mammograms more effectively.