Weakly and semi supervised detection in medical imaging via deep dual branch net

This study presents a novel deep learning architecture for multi-class classification and localization of abnormalities in medical imaging illustrated through experiments on mammograms. The proposed network combines two branches. One branch is region with newly added normal-region class. Second detection ranking regions relative to one another. Our method enables at full mammogram resolution both weakly semi-supervised settings. A objective function allows the incorporation local annotations into model. We present impact our schemes several performance measures localization, evaluate cost effectiveness lesion annotation effort. evaluation was primarily conducted over large multi-center mammography dataset ~3,000 mammograms various findings. results supervised showed significant improvement compared previous approaches. show that time consuming involved can be addressed by leverage subset locally annotated data. Weakly methods coupled produce effective explainable model adopted radiologists field.