MDS-UNet: Multi-Dimensional and Multi-Scale UNet for Medical Image Segmentation

Transformer and its variants, along with Mamba and its variants, have found extensive applications in medical image segmentation tasks. However, these hybrid models are often burdened with significant parameter counts and computational demands, which restrict their effectiveness in environments with limited resources. To address this issue, we propose a lightweight medical image segmentation network, MDS-UNet, which integrates the Multi-Scale Receptive Field Fusion Module, Multi-Axis Attention Module, and Multi-Scale Masked Feature Aggregation Module. This design enables the extraction of both global and local features from multiple dimensions and multiple scales, effectively capturing contextual information. The model achieves superior segmentation performance as a lightweight network. The proposed model was validated using the ISIC 2018 skin lesion dataset, and experimental results indicate that MDS-UNet achieves superior performance while substantially reducing the number of parameters to 0.77M and the computational complexity to 1.52 GFLOPs. Specifically, the model attained a mean Intersection over Union of 81.73% and a Dice Similarity Coefficient of 90.32%. Compared to existing advanced models, MDS-UNet exhibits notable advantages in both lightweight design and segmentation accuracy.