GMmorph: dynamic spatial matching registration model for 3D medical image based on gated Mamba

Area of Science:

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

Background:

• Deformable registration aligns medical images using dense displacement fields, crucial for surgical planning and navigation.
• Current deep learning methods face challenges with complex displacements, feature interaction, and spatial matching, impacting accuracy and robustness, especially with abnormal tissues.

Purpose of the Study:

• To develop a novel dual-branch interactive deformable registration model that enhances spatial position matching for improved accuracy and robustness.
• To address limitations in current deep learning-based registration methods concerning complex displacements and feature integration.

Main Methods:

• Proposed a dual-branch interactive registration architecture incorporating a dynamic matching module for learnable offsets and trilinear interpolation for flexible feature expression.
• Integrated a gated mamba layer for global pixel-level feature modeling and a detail enhancement module with channel and spatial attention for local feature enrichment.
• Employed implicit regularization via a consistency loss to balance accuracy and minimize folding, supporting unsupervised and semi-supervised learning modes.

Main Results:

• The model demonstrated state-of-the-art performance in both single-modal and multi-modal image registration tasks.
• Evaluated on diverse datasets including normal brain, brain tumor, and lung images, showcasing robust performance.
• Achieved precise registration across various medical data types, validating its effectiveness.

Conclusions:

• The novel perspective of position matching in the proposed model significantly enhances deformable image registration accuracy and robustness.
• The model offers substantial clinical value for diverse medical applications requiring precise image alignment.
• The developed architecture effectively integrates global and local features, overcoming limitations of existing deep learning registration techniques.