A Population-Based Hyperregression Model to Accelerate Instance-Specific Real-Time Volumetric MRI Estimation
Abstract
Purpose
Implicit neural representations (INRs) enable continuous, resolution-agnostic modeling of complex anatomical motion and have shown remarkable promise for deformation-driven, instance-specific real-time volumetric MRI estimation from a prior MRI. However, training INRs from scratch for each instance faces a trade-off between computational efficiency and representation capacity, limiting their deployment in time-sensitive clinical workflows. We develop a population-based HyperRegression model to accelerate instance-specific INR optimization for real-time MRI estimation (HR-INR).
Methods
HR-INR adopts a registration-driven framework where an INR predicts deformation-vector-fields (DVFs) to warp a fully sampled source volume to a real-time target, with instance-specific optimization enforced by k-space data fidelity between the warped source and the undersampled target. To reduce the heavy cost of iterative optimization arising from repeated image-to-k-space transformations, a HyperRegression hypernetwork is introduced to predict INR weights for new cases through population-level generalization, thereby accelerating instance-specific optimization. The pipeline consists of three stages. First, a set of INRs are optimized to register fully sampled source-target volume pairs across the training dataset. Subsequently, the HyperRegression model is trained in the INR weight space to learn a conditional mapping from source-target k-space data to instance-specific INR weights. At test time, the trained HyperRegression model predicts INR weights for unseen source-target k-space data, which are loaded into the INR for k-space-domain instance-specific optimization. By transferring population-level knowledge to INR initialization, the framework achieves faster instance-specific optimization for volumetric MRI estimation.
Results
Experiments on 11 4D-MRI datasets demonstrated that HR-INR enabled volumetric MRI estimation from highly-undersampled k-space data and achieved an average Dice coefficient of 0.876±0.065 and a center-of-mass-error of 1.54±1.06mm for cardiac structure tracking, while enabling efficient instance-specific optimization (268s vs. 1071s from scratch) on an NVIDIA H100 GPU.
Conclusion
HR-INR effectively accelerated instance-specific INR optimization, demonstrating strong potential for accurate real-time volumetric MRI estimation from extremely under-sampled k-space data.