Challenge-aware U-net for breast lesion segmentation in ultrasound images

Deep learning methods can enhance the efficiency of tumor segmentation in breast ultrasound (BUS) images. However, noise interference, small tumors, and blurred boundaries can reduce segmentation accuracy. We design a three-branch challenge-aware U-net (CAU-net) to address these main challenges in BUS images. Our CAU-net extracts the features from three challenge-aware encoders in parallel first. Secondly, we propose an adaptive aggregation layer (AAL) to merge the multi-scale features of three challenging branches, enabling the network to adaptively handle different breast lesion samples with these main challenges. To further enhance the accuracy of segmentation, we introduce the graph reasoning module (GRM) to the network to model the correlation between the channels of the features and acquire the global information in the features. The result of our experiment on two datasets demonstrates the superiority of CAU-net over the advanced medical image segmentation methods. Our code can be downloaded from https://github.com/tzz-ahu.