Multi-scheme cross-level attention embedded U-shape transformer for MRI semantic segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate segmentation of Magnetic Resonance Imaging (MRI) is vital for disease diagnosis.
• Current Transformer-based methods struggle with capturing fine details, leading to boundary inaccuracies.
• Existing models lack MRI-specific embedding strategies for attention modules, limiting performance.

Purpose of the Study:

• To propose a novel Transformer model, MSCL-SwinUNet, for enhanced MRI image segmentation.
• To address limitations in detail capture and MRI-specific feature representation in current methods.
• To improve the accuracy and generalizability of Transformer-based segmentation in medical imaging.

Main Methods:

• Developed the Multi-Scheme Cross-Level Attention embedded U-shape Transformer (MSCL-SwinUNet).
• Integrated cross-level spatial-wise attention (SW-Attention) for detailed information transfer.
• Incorporated cross-stage channel-wise attention (CW-Attention) and multi-stage scale-wise attention (ScaleW-Attention) for feature refinement.

Main Results:

• MSCL-SwinUNet demonstrated superior accuracy and generalizability on ACDC, MM-WHS, and Synapse datasets.
• The model effectively preserved detailed boundaries, outperforming state-of-the-art methods.
• Visualizations confirmed the model's capability in precise segmentation and localization.

Conclusions:

• MSCL-SwinUNet significantly advances Transformer-based segmentation for medical imaging applications.
• The proposed attention mechanisms offer new insights for designing MRI-specific embedding paradigms.
• This work enhances diagnostic capabilities through improved MRI segmentation accuracy.