ZR²ViM: a recursive vision Mamba model for boundary-preserving medical image segmentation

Area of Science:

• Medical image analysis
• Deep learning for medical imaging
• Computer vision

Background:

• Medical image segmentation is crucial for quantitative disease analysis and treatment planning.
• Existing deep learning methods face challenges in modeling long-range dependencies and maintaining boundary accuracy, especially for complex or low-contrast structures.
• Limited computational resources hinder the simultaneous optimization of these factors.

Purpose of the Study:

• To introduce ZR²ViM, a novel recursion-enhanced visual state space model for medical image segmentation.
• To address the limitations of current methods in handling complex morphologies and blurred edges.
• To improve both region consistency and boundary localization in medical image segmentation tasks.

Main Methods:

• Proposed ZR²ViM, augmenting the Vision Mamba framework with a Zigzag Recursive Reinforced (ZR²) Block.
• Incorporated Stacked State Redistribution (SSR) and Nested Recursive Connection (NRC) for iterative fusion of local details and global context.
• Utilized a Cross-directional Zigzag WKV (CZ-WKV) module with Quad-Directional Token Shift (Q-Shift) for multi-step recursive updates and spatial directional information injection.
• Achieved near-linear computational complexity.

Main Results:

• ZR²ViM demonstrated superior performance across four medical imaging domains (dermatoscopy, breast ultrasound, colorectal polyps, abdominal CT) on five public datasets.
• Consistently outperformed convolutional, attention-based, and other visual state space architectures in region consistency and boundary localization.
• Achieved a 2.15 mm reduction in HD95 on the Synapse multi-organ CT dataset compared to the CC-ViM baseline, indicating precise boundary delineation.

Conclusions:

• The ZR²ViM framework provides accurate, boundary-preserving medical image segmentation across various modalities and complex structures.
• Achieves these results with near-linear computational complexity, offering a robust and efficient solution.
• Establishes a promising foundation for advanced clinical and research applications in medical image analysis.