cMeta-INR: cohort-informed meta-learning-based implicit neural representation for deformable registration-driven real-time volumetric MRI estimation

<i>Objective.</i>Rapid and accurate reconstruction of high-quality three-dimensional magnetic resonance (MR) images from undersampled<i>k</i>-space data with variable sampling patterns remains a challenge due to limited available information and the need to preserve rich anatomical details. Deformable image registration provides a promising solution by warping a fully-sampled reference image to align with undersampled data acquired on-board (from an image-guided treatment delivery platform like MR-LINACs). In this study, we proposed a novel registration framework-cohort-informed meta-learning (cMeta)-to enhance the accuracy and efficiency of implicit neural representations (INR) for limited<i>k</i>-space data-driven patient-specific deformable image registration.<i>Approach.</i>cMeta-INR incorporated token-aware modulation and population-level deformation priors to guide an INR template-based meta-learning process. By encoding contextual information and leveraging cohort-informed priors, the resulting meta-learning framework enabled the INR to rapidly adapt to new registration cases with undersampled<i>k</i>-space data. Specifically, for the meta learning, a modulation module with token-awareness was introduced to modulate the INR template, and a pre-trained population-based registration network (KS-RegNet) was employed to generate coarse, reference deformation vector fields and latent embeddings for computing the deformation discrepancy loss and embedding similarity loss. During test-time adaptation, the INR, initialized from the meta-learned template, was efficiently fine-tuned to new<i>k</i>-space data with minimal iterations.<i>Main results.</i>Experiments were conducted on 14 abdominal and 11 cardiac 4D magnetic resonance imagings (MRIs) with 5-13 spokes. cMeta-INR outperformed state-of-the-art methods, achieving the best average (± s.d.) Dice similarity coefficients (0.778 ± 0.056 for abdominal and 0.894 ± 0.076 for cardiac data), and center-of-mass errors (3.04 ± 1.48 mm and 1.32 ± 1.02 mm, respectively), while enabling rapid test-time adaptation of only ∼35 s on an NVIDIA H100 GPU.<i>Significance.</i>The proposed cohort-informed meta-learning framework effectively enhanced the adaptation capabilities of INRs to individual patients under highly undersampled<i>k</i>-space scenarios, demonstrating strong potential for fast and accurate patient-specific deformable registration.