Self-supervised dynamic MRI reconstruction

Abstract

Deep learning techniques have recently been adopted for accelerating dynamic MRI acquisitions. Yet, common frameworks for model training rely on availability of large sets of fully-sampled MRI data to construct a ground-truth for the network output. This heavy reliance is undesirable as it is challenging to collect such large datasets in many applications, and even impossible for high spatiotemporal-resolution protocols. In this paper, we introduce self-supervised training to deep neural architectures for dynamic reconstruction of cardiac MRI. We hypothesize that, in the absence of ground-truth data, elevating complexity in self-supervised models can instead constrain model performance due to the deficiencies in training data. To test this working hypothesis, we adopt self-supervised learning on recent state-of-the-art deep models for dynamic MRI, with varying degrees of model complexity. Comparison of supervised and self-supervised variants of deep reconstruction models reveals that compact models have a remarkable advantage in reliability against performance loss in self-supervised settings.