Uncertainty-Aware Self-supervised Neural Network for Liver $T_{1ρ}$ Mapping with Relaxation Constraint
Authors:
Chaoxing Huang,
Yurui Qian,
Simon Chun Ho Yu,
Jian Hou,
Baiyan Jiang,
Queenie Chan,
Vincent Wai-Sun Wong,
Winnie Chiu-Wing Chu,
Weitian Chen
Abstract:
$T_{1ρ}$ mapping is a promising quantitative MRI technique for the non-invasive assessment of tissue properties. Learning-based approaches can map $T_{1ρ}$ from a reduced number of $T_{1ρ}$ weighted images, but requires significant amounts of high quality training data. Moreover, existing methods do not provide the confidence level of the $T_{1ρ}…
▽ More
$T_{1ρ}$ mapping is a promising quantitative MRI technique for the non-invasive assessment of tissue properties. Learning-based approaches can map $T_{1ρ}$ from a reduced number of $T_{1ρ}$ weighted images, but requires significant amounts of high quality training data. Moreover, existing methods do not provide the confidence level of the $T_{1ρ}$ estimation. To address these problems, we proposed a self-supervised learning neural network that learns a $T_{1ρ}$ mapping using the relaxation constraint in the learning process. Epistemic uncertainty and aleatoric uncertainty are modelled for the $T_{1ρ}$ quantification network to provide a Bayesian confidence estimation of the $T_{1ρ}$ mapping. The uncertainty estimation can also regularize the model to prevent it from learning imperfect data. We conducted experiments on $T_{1ρ}$ data collected from 52 patients with non-alcoholic fatty liver disease. The results showed that our method outperformed the existing methods for $T_{1ρ}$ quantification of the liver using as few as two $T_{1ρ}$-weighted images. Our uncertainty estimation provided a feasible way of modelling the confidence of the self-supervised learning based $T_{1ρ}$ estimation, which is consistent with the reality in liver $T_{1ρ}$ imaging.
△ Less
Submitted 25 October, 2022; v1 submitted 7 July, 2022;
originally announced July 2022.