IPOD:Inverse-Problem-Driven Meta-Learning for Fast Generalizable Neural Representations in MRI Reconstruction

ICLR 2026 Conference Submission18620 Authors

Keywords: meta-learning. implicit neural representation. MRI reconstruction

Abstract: Implicit neural representation (INR) demonstrates strong performance in magnetic resonance imaging (MRI) reconstructions by learning continuous mappings from spatial coordinates to signal intensities. However, existing unsupervised INR approaches require training from scratch for each observation, which is time-consuming and limits practical deployment. In this work, we propose an inverse-problem-driven meta-learning framework (iPod) that learns generalizable parameter initializations for INR directly from various undersampled reconstruction tasks without requiring fully sampled references. Technically, the meta-update is adaptively modulated by the hyperparameters performance of each inverse problem, ensuring optimal parameter distributions for robust and efficient initialization. Our approach leverages diverse reconstruction tasks with varying sampling patterns and anatomical structures to acquire a powerful and robust prior. Experimental validations demonstrate that the proposed framework provides powerful initialization that achieves fast convergence and superior reconstruction quality across different imaging protocols, outperforming baseline INR methods. Furthermore, this framework eliminates the dependence on reference images in conventional meta-learning procedures and has the potential to be extended to INR-based solutions for a wide range of imaging inverse problems. The code and data will be available at: https://anonymous.4open.science/r/iPod-2C60

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 18620