Uncertainty-aware Quantitative MRI Reconstruuction using conidtional Wasserstein GAN

Motivation: Recent advancements in deep learning hold great promise for expediting MRI scans, particularly in the time-consuming domain of quantitative MRI. However, the inherent "black-box" nature of deep learning models introduces an element of unknown risk when confronted with unseen data. This is a critical concern in qMRI, where the utmost reliability is imperative. Goal(s): To develop a qMRI reconstruction framework which could quantifies model uncertainty for guiding clinical decisions. Approach: We presented a conditional Wasserstein GAN for qMRI reconstruction, enabling uncertainty assessment through posterior sampling. Results: The proposed method achieved comparable performance to the current method while offering valuable pixel-wise uncertainty maps. Impact: The study offers clinicians and researchers a reliable qMRI reconstruction method with pixel-wise uncertainty assessment. This could spark further investigations into model reliability and potentially facilitate the practical application of deep learning-based qMRI methods.