An effective deep learning algorithm for medical image registration.

Area of Science:

• Medical Imaging
• Computational Anatomy
• Image Registration

Background:

• Accurate medical image registration is vital for applications like longitudinal monitoring and multimodal fusion.
• Ensuring topological consistency and invertibility during registration remains a significant challenge.
• Traditional methods may allow local folding, compromising anatomical integrity.

Purpose of the Study:

• To propose DTC-Reg, a dynamically learned registration framework explicitly enforcing diffeomorphic deformation.
• To address limitations of penalty-based regularization methods that may permit folding.
• To introduce a Multiscale Folding-aware Deformation Correction (MFDC) module for enhanced registration.

Main Methods:

• DTC-Reg integrates homotopy-based control-increment formulation with multiscale geometric constraints.
• Parameter-sharing U-Nets extract multiscale feature pyramids, feeding into a symmetric registration module.
• A sequential temporal cascade network refines deformation fields, enhanced by the MFDC module for folding correction.

Main Results:

• The proposed DTC-Reg framework successfully enforces diffeomorphic deformation, reducing folding.
• The MFDC module significantly improves deformation regularity and can be integrated into other registration networks.
• Experiments on 3D brain MRI data show superior quantitative and qualitative performance compared to existing methods.

Conclusions:

• DTC-Reg provides a robust solution for accurate and topologically consistent medical image registration.
• The MFDC module offers a valuable tool for improving the diffeomorphic consistency of registration algorithms.
• The framework demonstrates significant potential for advancing medical imaging analysis, particularly in neuroimaging.