L2NLF: a novel linear-to-nonlinear framework for multi-modal medical image registration

Area of Science:

• Medical imaging
• Deep learning
• Computer vision

Background:

• Multimodal medical image registration is complex.
• Deep learning enhances non-rigid registration accuracy.
• Existing methods face challenges in multimodal scenarios.

Purpose of the Study:

• To propose a novel linear-to-nonlinear framework (L2NLF) for multimodal medical image registration.
• To introduce a new deep neural network, CrossMorph, for unsupervised deformable registration.
• To evaluate the effectiveness and authenticity preservation of the proposed framework.

Main Methods:

• A two-stage framework: linear image conversion followed by nonlinear deep learning-based registration.
• Design of the CrossMorph network utilizing a U-net-like structure with CrossFormer blocks and Booster.
• Unsupervised learning approach for deformable registration.

Main Results:

• L2NLF achieves excellent registration with minimal computation and preserves image authenticity.
• CrossMorph outperforms state-of-the-art methods in reducing average surface distance and improving Dice scores.
• The deformation fields generated by CrossMorph exhibit enhanced smoothness.

Conclusions:

• The proposed L2NLF framework and CrossMorph network offer a robust solution for multimodal medical image registration.
• The method demonstrates significant improvements in accuracy and efficiency.
• The approach holds potential for valuable clinical applications in medical imaging.