FRM: A novel Faster-RCNN Mutant network for breast lesions screening

In recent years, the Faster-RCNN network is widely used for DICOM images pathological screening in medicine. In the breast field, the Faster-RCNN algorithm model was trained with mammogram image data to screen mass and calcification, respectively. Not only reducing the workload on doctors but also lifting efficiency and lowering patient costs. While Focal loss (FL) is a loss function that can keep the balancing of the learning ability between SHL (sample which is hard to learn) and SEL (sample which is easy to learn), and implemented the evolution from two-stage to one-stage named Retinanet based on the standard Faster-RCNN network for object detection. Although FL can enhance the generalization ability of CNN to SHL, the learning weight of each category is a non-sensitive parameter. Inspired by the FL function, to make the learning weight a sensitive parameter, we modified the Faster-RCNN network structure named the Faster-RCNN-Mutant and designed a mini-dataset dynamic weights (MDDW) loss function with a regularization item based on FL by calculating the distribution of SHL and SEL from the result of RPN network and to carry out the sensitivity of parameters. In this work, compared to the results of Faster-RCNN on the VOC2007 open-source dataset, F-SHL and FA-SHL were improved by 35.29% and 30%, respectively. The precision of benignCalcification and cancerMass improved by 0.48% and 0.73% on the DDSM dataset, but the benignMass has no changes the precision is 0%, and Faster-RCNN-Mutant ultimate recognition precision of cancerMass, benignMass, and benign-Calcification are 26.17%, 0.00%, and 9.68%, respectively.