Learning Deformable Image Registration From Optimization: Perspective, Modules, Bilevel Training and Beyond.

Area of Science:

• Medical image analysis
• Computational anatomy
• Machine learning in medical imaging

Background:

• Conventional deformable registration is computationally expensive and lacks speed.
• Current deep learning methods lack explicit geometric constraints, leading to implausible deformations.
• Hyper-parameter tuning in deep learning registration is resource-intensive.

Purpose of the Study:

• To develop a learning-based framework for efficient and geometrically constrained diffeomorphic image registration.
• To address limitations of existing deep learning approaches in medical image registration.
• To enable robust registration across diverse medical imaging datasets and modalities.

Main Methods:

• A generic optimization model for diffeomorphic registration.
• Learnable architectures for multi-scale, coarse-to-fine feature propagation.
• A bilevel self-tuned training strategy for efficient hyper-parameter optimization.

Main Results:

• State-of-the-art performance in 3D brain MRI and liver CT registration.
• Demonstrated diffeomorphic guarantee and extreme computational efficiency.
• Successful application to multi-modal image registration and downstream tasks like segmentation.

Conclusions:

• The proposed framework achieves efficient, accurate, and geometrically plausible medical image registration.
• The bilevel self-tuned training strategy enhances flexibility and reduces computational burden.
• This method supports advanced medical image analysis tasks, including fusion and segmentation.