Joint Supervised and Self-Supervised Learning for MRI Reconstruction

Abstract

Magnetic Resonance Imaging (MRI) is a crucial modality but, its inherently slow acquisition process poses challenges in obtaining fully-sampled $k$-space data under motion. The lack of fully-sampled acquisitions, serving as ground truths, complicates the training of deep learning (DL) algorithms in a supervised manner. To address this limitation, self-supervised learning (SSL) methods have emerged as a viable alternative, leveraging available subsampled $k$-space data to train neural networks for MRI reconstruction. Nevertheless, these approaches often fall short when compared to supervised learning (SL). We propose Joint Supervised and Self-supervised Learning (JSSL), a novel training approach for DL-based MRI reconstruction algorithms aimed at enhancing reconstruction quality in cases where target datasets containing fully-sampled $k$-space measurements are unavailable. JSSL operates by simultaneously training a model in a SSL setting, using subsampled data from the target dataset(s), and in a SL manner, utilizing proxy datasets with fully-sampled $k$-space data. We demonstrate JSSL's efficacy using two distinct combinations of target and proxy data. Quantitative and qualitative results showcase substantial improvements over conventional SSL methods. Furthermore, we provide "rule-of-thumb" guidelines for training MRI reconstruction models. Our code is available at https://github.com/NKI-AI/direct.