Speaker
Description
Data-driven reconstruction methods, mostly based on deep learning, have emerged as the undisputed state-of-the-art across many different inverse problems, including MRI reconstruction. However, they are also notorious for requiring large and diverse datasets to be trained successfully, preferably in a supervised manner. When target reconstructions cannot be obtained, zero-shot self-supervised learning approaches have been proposed as adequate alternatives for model training, often requiring some additional mechanisms to avoid overfitting. In this work, we consider a zero-shot self-supervised training approach for a hybrid reconstruction method that is based in the combination of a hand-crafted regulariser and a deep neural network that adapts, in a spatio-temporal manner, the strength of the employed prior. We demonstrate that this combination allows for fast learning of the regularisation strength, achieving nearly the same performance as supervised pre-training combined with sample-wise, test-time self-supervised training. Further, we show that the combination of a hand-crafted prior with a learned adaptive strength systematically avoids overfitting, eliminating the need for strategies such as monitoring metrics on validation sets or early stopping. Finally, the considered methods are highly interpretable due to the choice of the hand-crafted priors. We showcase the proposed paradigm for spatio-temporally adaptive Total Variation (TV) and Total Generalised Variation (TGV) applied to image denoising and dynamic cardiac MRI reconstruction.