UNIMODAL CYCLIC REGULARIZATION FOR TRAINING MULTIMODAL IMAGE REGISTRATION NETWORKS.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Unsupervised multimodal image registration is crucial for integrating information from different imaging modalities.
• Current deep learning approaches excel at learning similarity metrics but often rely on hand-crafted regularization terms.
• This reliance on artificial constraints limits the adaptability and performance of multimodal registration frameworks.

Purpose of the Study:

• To develop an automated regularization method for unsupervised multimodal image registration.
• To leverage task-specific prior knowledge from unimodal registration to enhance multimodal registration accuracy.
• To improve the estimation of deformation fields in challenging registration scenarios.

Main Methods:

• Proposed a unimodal cyclic regularization training pipeline.
• Learned task-specific prior knowledge from simpler unimodal registration tasks.
• Applied the learned priors to constrain the deformation field in multimodal registration.

Main Results:

• The proposed cyclic regularization method demonstrated superior performance compared to conventional regularization techniques.
• Significant improvements were observed in abdominal CT-MR registration experiments.
• The method was particularly effective in handling severely deformed local regions.

Conclusions:

• The learned unimodal cyclic regularization effectively constrains multimodal registration deformation fields.
• This approach offers a more adaptive and data-driven alternative to hand-crafted regularization.
• The findings suggest a promising direction for advancing unsupervised multimodal image registration.