Add to Chrome
Add to Firefox
Add to EdgeDeep Image prior with StruCtUred Sparsity (DISCUS) for dynamic MRI reconstruction
Dec 01, 2023
High-quality training data are not always available in dynamic MRI. To address this, we propose a self-supervised deep learning method called deep image prior with structured sparsity (DISCUS) for reconstructing dynamic images. DISCUS is inspired by deep image prior (DIP) and recovers a series of images through joint optimization of network parameters and input code vectors. However, DISCUS additionally encourages group sparsity on frame-specific code vectors to discover the low-dimensional manifold that describes temporal variations across frames. Compared to prior work on manifold learning, DISCUS does not require specifying the manifold dimensionality. We validate DISCUS using three numerical studies. In the first study, we simulate a dynamic Shepp-Logan phantom with frames undergoing random rotations, translations, or both, and demonstrate that DISCUS can discover the dimensionality of the underlying manifold. In the second study, we use data from a realistic late gadolinium enhancement (LGE) phantom to compare DISCUS with compressed sensing (CS) and DIP and to demonstrate the positive impact of group sparsity. In the third study, we use retrospectively undersampled single-shot LGE data from five patients to compare DISCUS with CS reconstructions. The results from these studies demonstrate that DISCUS outperforms CS and DIP and that enforcing group sparsity on the code vectors helps discover true manifold dimensionality and provides additional performance gain.
