RMDA-Net: A 3D Residual Multi-Scale Dynamic Attention Network for Small-Structure Knee MRI Segmentation

Knee joint magnetic resonance imaging (MRI) segmentation plays a vital role in the early diagnosis and quantitative analysis of knee osteoarthritis. However, the task remains challenging due to the complex anatomical structures, blurred boundaries, and the presence of small structures of cartilages. To address these challenges, we propose a robust 3D Residual Multi-Scale Dynamic Attention Network (RMDA-Net) for automated segmentation of knee cartilage and bone. Specifically, we design a Residual Multi-Scale Dilated Depthwise Convolution (RMDDW) module to capture the local and global context efficiently, and introduce a Dynamic Gate Attention (DGA) mechanism to adaptively fuse features from skip connections, which together improve the feature representation of complex anatomical structures and enhance boundary delineation. In addition, a Multi-Scale Auxiliary Output (MAO) with a boundary-aware loss is integrated to strengthen the boundary localization across all structures, particularly benefiting small tissues. Experiments on the public OAI-ZIB dataset demonstrate that RMDA-Net achieves a mean Dice score of 93.87%, an IoU of 88.75%, and an ASSD of 0.115 mm, outperforming state-of-the-art methods while maintaining reasonable computational complexity. Additional experiments on the independent SKI10 dataset further demonstrate the robustness and generalization capability of the proposed method across different imaging distributions.