An Efficient Lightweight Network for Breast Cancer Segmentation

Image segmentation is widely used in breast can-cer treatment by rapidly localizing lesions. Most segmentation strategies are both time-consuming and labor-intensive, creating an urgent need for an automatic lesion segmentation algorithm to assist ultrasound physicians. CNN-based computer-aided segmentation methods have achieved remarkable effects, particularly with the U-Net series of networks. However, existing methods improve segmentation accuracy by increasing compu-tational complexity, which creates challenges for deployment on ultrasound examination machines with limited computational re-sources. The proposed Convolutional Block Attention Distillation U-Net (CBADUNet) in this paper integrates a Convolutional Block Attention Distillation Block (CBADB) and employs multi-branch deep supervision (MBDS). CBADB integrates attention mechanisms and feature distillation to extract discriminative semantic information from low-frequency features during the encoder stage. This ensures that the nuances and details captured are accurately represented and integrated during the forward propagation, enhancing the overall fidelity and precision of the output. Meanwhile, MBDS utilizes multiple auxiliary branches to adaptively integrate features both the encoder and decoder. These auxiliary branches, which are discarded during the inference stage, effectively facilitate information fusion without increasing parameters. The results from the Breast Ultrasound Imaging (BUS I) dataset indicate that the proposed CBADUNet provides a better trade-off compared to advanced medical image segmen-tation methods, balancing enhanced performance with reduced model complexity.