Multi-task Magnetic Resonance Imaging Reconstruction using Meta-learning
Wanyu Bian, Albert Jang, Fang Liu
TL;DR
The paper addresses the challenge of reconstructing MR images when data come from multiple imaging sequences with different contrasts. It introduces a bi-level meta-learning framework that combines an unrolled proximal gradient descent for cross-domain learning (image and k-space) with a meta-learner that captures and distributes shared knowledge across tasks. The approach, MTML, includes base-learners for per-task learning and a meta-learner Z that channels meta-knowledge Θ to task reconstructions, optimized via a MAML-style training procedure. Experiments on multi-contrast knee MRI demonstrate that MTML outperforms single-task baselines, improving PSNR, SSIM, and NMSE while preserving texture and edge details, suggesting robust, efficient multi-contrast reconstruction with potential clinical impact.
Abstract
Using single-task deep learning methods to reconstruct Magnetic Resonance Imaging (MRI) data acquired with different imaging sequences is inherently challenging. The trained deep learning model typically lacks generalizability, and the dissimilarity among image datasets with different types of contrast leads to suboptimal learning performance. This paper proposes a meta-learning approach to efficiently learn image features from multiple MR image datasets. Our algorithm can perform multi-task learning to simultaneously reconstruct MR images acquired using different imaging sequences with different image contrasts. The experiment results demonstrate the ability of our new meta-learning reconstruction method to successfully reconstruct highly-undersampled k-space data from multiple MRI datasets simultaneously, outperforming other compelling reconstruction methods previously developed for single-task learning.