Dual-Path Framework for Intra-Class Imbalance Medical Image Segmentation

The intra-class imbalance usually occurs in medical images due to external influences, such as noise interference and changes in camera angle. It leads to complex textures and varied appearances within the target object region and makes segmentation task challenging. To deal with this kind of problem, we proposed a dual-path framework in this paper. Considering that the object consists of two subclasses (majority- and minority-subclass), a deep learning model is adopted to separate them. We constructed two weighted maps for the dual paths, related to majority- and minority-subclass respectively. A fusion module was designed to generate the final output according to the results from the dual paths. The experimental results on two datasets shew our approach's validity and superiority for medical image segmentation compared with other competing methods.