ECU-Net: multi-scale salient boundary detection and contrast feature enhancement U-Net for breast ultrasound image segmentation

U-Net and its variants show excellent performance in medical image segmentation. However, due to the information loss caused by down-sampling and the blurry boundary of the ultrasound images itself, it is hard to extract and recover boundary information. Additionally, due to the intensity similarity and low contrast between tumor and non-tumor regions in breast ultrasound images, obtaining discriminative features is difficult. So we propose an optimized U-Net breast ultrasound image segmentation model ECU-Net. Firstly, a multi-scale salient boundary detection module is proposed to enhance the model’s perception of tumor boundaries. Secondly, a contrast feature enhancement module is proposed to highlight the otherness between tumor and normal tissues. Then we combine the two modules to optimize skip-connection through boundaries completion and contrast information enhancement. Exhaustive experiments are conducted on two publicly available datasets, and experimental results show our ECU-Net performs better than U-Net and its classic variants.