AMCU-Net: Adaptive Multi-Scale ConvMixer U-Net for Breast Ultrasound Lesion Segmentation

Breast cancer is the most common malignant tumor affecting women globally, with early detection crucial to improving patient outcomes. Traditional segmentation methods, while proficient in global feature extraction, often fail to capture fine-grained local features across varying scales, resulting in missed detections and inaccurate segmentation of small lesions. To tackle these challenges, a network architecture, AMCU-Net, is proposed, which is designed to enhance segmentation accuracy for lesions of diverse scales and shapes. AMCU-Net adopts a multi-module structure to improve sensitivity to lesions at different scales, thereby enhancing small-lesion capture and optimizing segmentation across multiple resolutions. Furthermore, AMCU-Net integrates Convolutional Block Attention Module and Multi-Scale Attention Gate, dynamically focusing on essential spatial and channel features to bolster feature extraction and ensure effective small-lesion delineation. This approach enables precise boundary segmentation and high accuracy under various resolutions. Additionally, the ConvMixer module is employed to effectively merge multi-scale features through mixed convolution operations, boosting the model's capability to accurately segment complex heterogeneous lesions and adapt to images of different sizes. Experimental results demonstrate that AMCU-Net excels in multi-scale segmentation tasks, significantly advancing small-lesion delineation and overall segmentation performance.