Instabilities in Conventional Multi-Coil MRI Reconstruction with Small Adversarial Perturbations

Although deep learning (DL) has recently received significant attention in accelerated MRI, recent studies suggest that small perturbations may lead to large instabilities in DL-based reconstructions. This has also highlighted concerns for their utility in clinical settings. However, these works focus on single-coil acquisitions, which are not practically relevant. In this work, we investigate how small adversarial perturbations affect multi-coil MRI reconstruction, particularly using conventional non-DL methods. Our results indicate that for multi-coil MRI reconstruction, conventional parallel imaging and multi-coil compressed sensing (CS) methods also exhibit considerable instabilities against small adversarial perturbations. Moreover, for physics-guided DL reconstructions that utilize the forward encoding operator explicitly, such small perturbations predominantly target the linear data-consistency units. These results suggest that at high acceleration rates, adversarial attacks exploit the ill-conditioning of the forward encoding operator.