AMDA-Net: Adaptive Multi-scale Deformable Attention U-Net for Medical Image Segmentation

In a variety of clinical applications, medical image segmentation is essential for accurately identifying and localizing lesions. Nevertheless, traditional methods frequently encounter problems like intra-class inconsistency and inter-class indistinction, especially when handling complex biological structures. To address these issues, this work proposes the AMDA-Net, a multi-scale adaptive deformable attention U-Net designed especially for medical image segmentation. Integrating the dynamic components of deformable convolution and incorporating an enhanced hybrid spatial attention module, the AMDA-Net enhances the model’s ability to acquire rich semantic features and multi-level contextual bonds. Experiments show that AMDA-Net outperforms current approaches, with an average Dice Similarity Coefficient of 84.78% on the Synapse multi-organ segmentation dataset. The presented methodology not only augments segmentation precision but also exhibits the efficacy and generality of the presented function.