MEM-UNet: Morphology-Enhanced 3D Mamba UNet for Esophagus Segmentation

Area of Science:

• Medical image analysis
• Deep learning for medical imaging
• Computational anatomy

Background:

• Accurate segmentation of irregular, low-contrast medical structures like the esophagus is challenging.
• Existing methods struggle with precise boundary delineation in 3D CT volumes.
• Need for advanced deep learning frameworks to enhance segmentation performance.

Purpose of the Study:

• To propose MEM-UNet, a 3D Mamba-based UNet framework enhanced with mathematical morphology for robust medical image segmentation.
• To improve the segmentation accuracy of challenging anatomical structures, particularly the esophagus.
• To leverage shape-awareness from morphological operations for enhanced boundary precision.

Main Methods:

• Developed a 3D Mamba backbone adapting State Space Models (SSM) for CT volumes.
• Integrated Morphology-Aware Spatial-Channel Attention (MASCA) blocks with Morphology-Enhanced Spatial Convolution (MESC) and Squeeze-and-Excitation (SE) in skip connections.
• Introduced a Morphology-Enhanced Decision (MED) layer for contour refinement and precise voxel-level classification.

Main Results:

• MEM-UNet achieved superior performance on SegTHOR and BTCV datasets compared to state-of-the-art models.
• Achieved Dice Similarity Coefficient (DSC) scores of 87.42% (multi-organ) and 78.94% (esophagus) on SegTHOR.
• Achieved DSC scores of 74.86% (multi-organ) and 67.70% (esophagus) on BTCV.
• Ablation studies validated the effectiveness of individual components and the overall pipeline.

Conclusions:

• MEM-UNet offers a robust and effective framework for 3D medical image segmentation, particularly for challenging structures.
• The integration of Mamba and mathematical morphology significantly enhances segmentation accuracy and boundary delineation.
• The proposed method represents a significant advancement in automated medical image segmentation technology.