cMeta-INR: cohort-informed meta-learning-based implicit neural representation for deformable registration-driven...

Area of Science:

• Medical Imaging
• Machine Learning
• Computational Anatomy

Background:

• Reconstructing high-quality 3D MR images from undersampled k-space data is challenging.
• Deformable image registration is crucial for aligning medical images, especially in image-guided treatments.
• Implicit Neural Representations (INRs) offer potential for detailed image reconstruction but struggle with limited data.

Purpose of the Study:

• To develop a novel cohort-informed meta-learning (cMeta-INR) framework.
• To enhance the accuracy and efficiency of INRs for patient-specific deformable image registration using limited k-space data.
• To enable rapid adaptation of INRs for undersampled MR imaging scenarios.

Main Methods:

• Proposed cMeta-INR framework incorporating token-aware modulation and population-level deformation priors.
• Utilized a pre-trained population-based registration network (KS-RegNet) for meta-learning.
• Employed INR template-based meta-learning for rapid adaptation to new registration cases with undersampled k-space data.

Main Results:

• cMeta-INR outperformed state-of-the-art methods on abdominal and cardiac 4D MRIs.
• Achieved superior Dice similarity coefficients (0.778 for abdominal, 0.894 for cardiac) and center-of-mass errors.
• Demonstrated rapid test-time adaptation in approximately 35 seconds on an NVIDIA H100 GPU.

Conclusions:

• The cohort-informed meta-learning framework significantly improves INR adaptation for individual patients with highly undersampled k-space data.
• cMeta-INR shows strong potential for fast and accurate patient-specific deformable registration in medical imaging.
• This approach addresses key challenges in real-time image-guided interventions and MR image reconstruction.