Learning Multi-Level Features for Breast Mass Detection
Qinggong Zeng, Huiqin Jiang, Ling Ma · 2018
In order to quickly detect masses from mammography images for the early screening of breast cancer, this paper proposes a breast mass detection improved algorithm based on Faster R-CNN. Firstly, we connect multi- level feature maps (conv-4, conv-5) in ZF model to generated candidate regions in RPN, then use the ROI pool layer to extract the features of the candidate regions. Finally the full connection layer output the region's classification score and the bounding box after regression. Experiments show that the detection sensitivity of this model for breast masses is 93.6%, and the average number of false positives per image is reduced to 0.651. Compared with the original model, the sensitivity of this one increases by 8.5 percentage points and its performance is excellent in the detection of breast masses.