HyperMorph

HyperMorph employs an amortized hyperparameter learning strategy using a hypernetwork to predict registration network parameters for learning-based deformable image registration, enabling efficient test-time hyperparameter exploration and generation of deformation fields without retraining.

Key Features:

• Efficient Hyperparameter Search: Facilitates fast, high-resolution hyperparameter searches at test time by conditioning registration network parameters on input hyperparameters.
• Enhanced Robustness: Improves robustness to model initialization by amortizing hyperparameter effects across a learned parameter mapping.
• Dataset-Specific Optimization: Rapidly identifies optimal hyperparameter values tailored to specific datasets, image contrasts, tasks, or anatomical regions.
• Single-Model Flexibility: Uses a single hypernetwork-trained model to produce parameters for different hyperparameter values, eliminating the need to retrain multiple registration models.

Scientific Applications:

• Deformable Image Registration: Produces deformation fields for learning-based deformable image registration applications requiring precise spatial correspondences in medical imaging.
• Task- and Contrast-Specific Tuning: Enables dataset- and task-specific hyperparameter tuning across different image contrasts and anatomical regions.

Methodology:

Trains a hypernetwork (meta network) via amortized hyperparameter learning to predict registration network parameters conditioned on input hyperparameters, enabling generation of deformation fields at test time without iterative pair-wise optimization or retraining.