Using Aleatoric Uncertainty to Aid Deep Learning based T1rho Mapping and Analysis in the Liver

Motivation: The utility of uncertainty to ensure a reliable learning-based parametric mapping in quantitative MRI is underexplored. Goal(s): This study aimed to develop a reliable method for quantitative T1rho mapping of liver using uncertainty-based deep learning. Approach: We proposed a parametric map refinement approach that trained the model probabilistically to estimate uncertainty in predicted T1rho values. The uncertainty map was used to enhance mapping performance and identify unreliable values in the region of interest. Results: Testing on 51 patients with liver fibrosis showed a mapping error of less than 3% and simultaneous uncertainty estimation. Impact: Our work demonstrates potential of saving scan time while preserving T1rho quantification accuracy. It is also shown that incorporating uncertainty estimation in the T1rho mapping network can improve the reliability of predicted values.