M³-TransUNet: Medical Image Segmentation Based on Spatial Prior Attention and Multi-Scale Gating.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Medical image segmentation faces challenges with complex anatomical structures at various scales.
• Current Transformer models struggle with multi-scale features, spatial modeling, and feature fusion.

Purpose of the Study:

• To propose M³-TransUNet, an architecture designed to overcome limitations in medical image segmentation.
• To enhance multi-scale representation, spatial modeling, and feature fusion in Transformer-based segmentation models.

Main Methods:

• Developed M³-TransUNet incorporating Multi-Scale Gate Attention (MSGA), Multi-Scale Selective Attention (MSSA), and Manhattan Enhanced Multi-Head Self-Attention (ME-MSA).
• Integrated Multi-kernel Gated Attention Gate (MKGAG) to refine skip connections, focusing on noise filtering and boundary detail preservation.
• Evaluated performance on public datasets: Synapse, CVC-ClinicDB, and ISIC.

Main Results:

• M³-TransUNet achieved state-of-the-art performance across multiple datasets.
• On the Synapse dataset, M³-TransUNet improved average Dice Similarity Coefficient (DSC) to 82.79% and reduced Hausdorff Distance 95th percentile (HD95) from 19.74 mm to 10.21 mm.
• Outperformed recent TransUNet variants like J-CAPA in segmentation accuracy and boundary detail.

Conclusions:

• M³-TransUNet effectively addresses multi-scale representation, spatial modeling, and feature fusion challenges in medical image segmentation.
• The proposed architecture demonstrates superior performance and potential for clinical applications requiring precise anatomical segmentation.