DHR-Net: Dynamic Harmonized registration network for multimodal medical images

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Anatomy

Background:

• Deep learning significantly advanced medical image registration, particularly in single-modality tasks.
• Multi-modal medical image registration faces challenges like anatomical inconsistency and unstable optimization.

Purpose of the Study:

• To propose an end-to-end framework, Dynamic Harmonized Registration framework (DHR-Net), for accurate multi-modal medical image registration.
• To enhance anatomical consistency and deformation field optimization in cross-modal registration.

Main Methods:

• DHR-Net utilizes a cascaded two-stage architecture with sequential translation and registration networks.
• A novel loss function based on Noise Contrastive Estimation maximizes mutual information for cross-modal consistency.
• Dynamic temperature adjustment optimizes feature contrast and preserves high-frequency details for topological structure preservation.

Main Results:

• DHR-Net demonstrated superior registration accuracy, deformation field smoothness, and cross-modal robustness compared to existing methods.
• The framework significantly improved cardiac image registration quality.
• Exceptional performance in preserving anatomical structures was observed.

Conclusions:

• DHR-Net offers a robust solution for multi-modal medical image registration challenges.
• The proposed method shows significant potential for clinical applications in medical imaging.