Deep implicit optimization enables robust learnable features for deformable image registration.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Deep Learning in Image Registration (DLIR) offers speed and weak supervision but lacks optimization benefits and inductive bias.
• Existing DLIR methods struggle with domain shift, leading to suboptimal performance.

Purpose of the Study:

• To bridge the gap between deep learning and optimization in image registration.
• To develop a DLIR method that incorporates optimization as a network layer for improved robustness and flexibility.

Main Methods:

• A deep network predicts multi-scale dense features, registered via an iterative optimization solver.
• The framework implicitly differentiates through the optimization solver, learning registration and label-aware features.
• Ensures warp functions are local minima of the registration objective in feature space.

Main Results:

• Achieves excellent performance on in-domain datasets and demonstrates robustness to domain shifts like anisotropy and varying intensity profiles.
• Enables switching between arbitrary transformation representations (e.g., free-form to diffeomorphic) at test time without retraining.
• Facilitates end-to-end feature learning for interpretability and arbitrary test-time regularization.

Conclusions:

• The proposed method combines the strengths of deep learning and optimization for superior image registration.
• Offers unprecedented flexibility in transformation representation and regularization at test time.
• Represents a significant advancement in DLIR, addressing limitations of existing approaches.