ARDMR: Adaptive recursive inference and representation disentanglement for multimodal large deformation registration

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Multimodal medical image registration is vital for liver cancer diagnosis and treatment.
• Challenges include significant intensity variations and large tissue deformations between different imaging modalities.
• Existing learning-based models struggle with these multimodal registration complexities.

Purpose of the Study:

• To develop an advanced multimodal deformable registration model (ARDMR) addressing modality discrepancies and large deformations.
• To improve the accuracy and robustness of medical image registration for liver cancer applications.
• To enhance the generalization capability of registration models on diverse datasets.

Main Methods:

• Proposed ARDMR model incorporating representation disentanglement with Multi-layer Contrastive Loss (MCL) for modality-invariant features.
• Introduced Multi-Scale Feature Registration (MSFR) module to handle complex deformations by integrating multi-scale features and deformation fields.
• Implemented an adaptive recursive inference strategy to optimize registration scale and iterations based on real-time performance.

Main Results:

• ARDMR demonstrated superior performance over state-of-the-art methods in qualitative and quantitative evaluations.
• Achieved a 2.5%-5% improvement in the Dice Similarity Coefficient (DSC) metric compared to the VoxelMorph baseline.
• Exhibited excellent robustness and generalization on distribution-shifted data.

Conclusions:

• The proposed ARDMR model effectively mitigates modality discrepancies and handles large deformations in multimodal medical image registration.
• ARDMR offers significant improvements in registration accuracy and robustness, particularly valuable for liver cancer analysis.
• The adaptive recursive inference strategy enhances the model's practical applicability and performance.