Boundary-enhanced sparse transformer for generalizable and accurate medical image segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Medical image segmentation is vital for computer-aided diagnosis, aiding in organ analysis, lesion detection, and treatment planning.
• Existing Transformer-based segmentation models struggle with high computational costs due to redundant global attention and poor reconstruction of fine-grained boundary details.

Purpose of the Study:

• To develop an efficient and accurate framework for general medical image segmentation.
• To address the limitations of Transformer-based models in handling redundant connections and reconstructing precise segmentation boundaries.

Main Methods:

• An encoder incorporating a frequency-domain similarity measure and a Key-Semantic Dictionary (KSD) to create sparse attention matrices, reducing redundancy and enhancing semantic relevance.
• A decoder featuring a learnable gradient-based operator to inject boundary-aware bias, improving the recovery of structural details along object contours.

Main Results:

• Achieved significant improvements in Dice scores and reductions in Hausdorff Distance (HD) across multiple datasets (ACDC, ISIC 2018, Synapse) compared to state-of-the-art baselines.
• Demonstrated substantial parameter reduction (up to 88.8% fewer parameters) compared to typical Transformer models, indicating enhanced efficiency.
• Showcased robustness across different medical imaging modalities, highlighting clinical applicability.

Conclusions:

• The proposed framework effectively balances segmentation accuracy, boundary fidelity, and computational efficiency.
• Frequency-domain sparse attention and learnable edge-guided decoding are key innovations for suppressing redundant correlations and enhancing structural detail reconstruction.
• The framework offers a lightweight, clinically applicable solution for high-precision medical image segmentation.