MSF-TransUNet: A Multi-Scale Feature Fusion Transformer-Based U-Net for Medical Image Segmentation with Uniform Attention

Abstract

Accurate medical image segmentation is essential for computer-assisted diagnosis and treatment systems. While conventional U-Net architectures and hybrid models integrating U-Net with Transformer networks have demonstrated remarkable performance in automatic segmentation tasks, these approaches frequently face challenges in effectively integrating multi-scale features. Additionally, semantic inconsistencies arising from simple skip connections during the encoding-decoding process remain problematic. To address these limitations, a novel architecture, MSF-TransUNet, is proposed, which incorporates a Feature Fusion Attention Block (FFA-Block) to enhance the fusion of multi-scale features. This approach facilitates dense feature interactions through the integration of uniform attention, achieving this with minimal computational overhead. The experimental results on the Synapse and ACDC medical image segmentation datasets reveal that MSF-TransUNet outperforms existing models. Specifically, the average Hausdorff Distance (HD) on the Synapse dataset is reduced to 22.40 mm, accompanied by an impressive Dice Similarity Coefficient (DSC) of 80.78%. Furthermore, the model achieves a DSC of 91.52% on the ACDC dataset, demonstrating its superior performance. These findings highlight the potential of MSF-TransUNet in advancing medical image segmentation by effectively addressing the challenges of multi-scale feature fusion and semantic consistency.