Enhancing medical image segmentation with MA-UNet: a multi-scale attention framework

Medical image segmentation is crucial for accurate diagnosis and treatment planning. Traditional methods struggle with complex medical images, while recent deep learning advancements, particularly U-Net and its variants, often suffer from insufficient feature extraction and misalignment issues. This paper introduces MA-UNet, an enhanced segmentation network based on U-Net, which integrates multi-scale features and a hybrid attention mechanism. MA-UNet is designed with a two-stage encoder comprising rough ordinary extraction and multi-scale fine extraction to improve feature representation. A hybrid feature optimization module is embedded in skip connections to address feature misalignment. Additionally, the Transformer in the bottleneck layer is optimized, and multi-scale convolution is utilized in the decoder. Experimental results on three datasets demonstrate that MA-UNet achieves optimal or near-optimal segmentation performance, outperforming ten state-of-the-art methods. The proposed framework significantly enhances medical image segmentation accuracy, potentially aiding clinicians in making more precise decisions. Our code is available at https://github.com/HZ-LL/MA-UNet.