Learning a Metric for Multimodal Medical Image Registration without Supervision Based on Cycle Constraints.

Area of Science:

• Medical imaging
• Deep learning
• Image registration

Background:

• Deep learning for medical image registration struggles with large transformations and limited supervision.
• Unsupervised, metric-based registration networks lack universal similarity metrics for multimodal data.
• Existing methods often require trade-offs between local and global feature metrics.

Purpose of the Study:

• To improve unsupervised, metric-based deep learning for medical image registration.
• To develop a method that overcomes limitations of handcrafted metric-based losses.
• To achieve robust registration for large transformations, including rigid alignment, in multimodal medical imaging.

Main Methods:

• Proposed a novel approach using synthetic three-way (triangular) cycles for unsupervised learning.
• Incorporated two multimodal transformations and one known synthetic monomodal transform within cycles.
• Developed a differentiable method for estimating large rigid transformations in end-to-end learning.

Main Results:

• Successfully tackled intra-patient abdominal CT-MRI registration.
• Achieved performance on par with state-of-the-art metric-supervision and classic registration methods.
• Demonstrated that cyclic constraints enable learning of cross-modality features for accurate anatomical alignment.

Conclusions:

• The proposed cyclic constraint method enhances deep learning-based medical image registration, particularly for large transformations.
• This approach provides a robust and effective solution for multimodal registration challenges, such as CT-MRI alignment.
• The method offers a promising alternative to handcrafted losses, improving accuracy and applicability in clinical settings.