LADDA: Latent Diffusion-Based Domain-Adaptive Feature Disentangling for Unsupervised Multi-Modal Medical Image...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Deformable image registration (DIR) is crucial for clinical diagnosis and treatment planning.
• Multi-modal registration faces challenges due to patient movement, intensity variations, and large deformations, leading to entangled anatomical and style information.
• Existing methods struggle with accurate anatomical alignment in the presence of these challenges.

Purpose of the Study:

• To propose a novel unsupervised framework for multi-modal medical image registration that addresses representation disentanglement.
• To improve the accuracy of anatomical alignment in multi-modal medical images by separating content and style information.
• To enhance the performance of medical image registration in challenging clinical scenarios.

Main Methods:

• Introduced the LAtent Diffusion-based Domain-Adaptive feature disentangling (LADDA) framework.
• Utilized a Latent Diffusion Model (LDM) to extract anatomical priors.
• Employed a Domain-Adaptive Feature Disentangling (DAFD) module to separate content and style features.
• Incorporated Neighborhood-Preserving Hashing (NPH) for hierarchical content integration and structural consistency.
• Developed a Unilateral-Query-Frozen Attention (UQFA) module to enhance feature coupling.

Main Results:

• The LADDA framework demonstrated superior performance in unsupervised multi-modal medical image registration.
• The method effectively disentangled anatomical and modality-specific style information.
• Achieved accurate anatomical alignment and improved recovery of detailed textures.
• Showcased strong generalization capabilities on out-of-distribution data across diverse clinical scenarios.

Conclusions:

• The proposed LADDA framework offers a robust solution for unsupervised multi-modal medical image registration.
• Explicitly addressing feature disentanglement significantly improves registration accuracy and robustness.
• The framework shows promise for enhancing clinical diagnosis and treatment planning through precise anatomical alignment.