Retrospective relaxometry from conventional contrasts by physics-informed deep learning: A pilot on Tumor, MS, Stroke and Epilepsy patients

Motivation: Quantitative MRI could enhance disease diagnosis, but current limited availability restricts its clinical application. Existing deep learning (DL) models for retrospective quantitative mapping lack evaluation of broader clinical populations. Goal(s): Evaluate the potential of retrospectively generated quantitative maps by physics-informed DL in healthy controls and a mixed-clinical population. Approach: Quantitative T1- and T2-maps were generated by a self-supervised, physics-informed DL model. The model was trained and evaluated on a heterogenous dataset, including tumor, multiple sclerosis, stroke, and epilepsy patients. Results: Generated quantitative maps agree with literature values for normal tissue and effectively discerned various lesions, demonstrating the potential of the model's generalization ability. Impact: This study demonstrates the use of self-supervised, physics-informed deep learning to generate full-brain quantitative T1- and T2-maps from conventional MRI on a heterogeneous dataset of neurological patients, potentially enabling future applications on large-scale datasets to improve diagnostic tools.