A Physics-Informed Convolutional Neural Network to Estimate Intravoxel Incoherent Motion Parameters in the Liver

Motivation: Intravoxel incoherent motion (IVIM) MRI produces diffusion estimates related to stages of liver fibrosis, however, there is an overlap of values between fibrosis stages and poor repeatability. Goal(s): We aimed to improve repeatability using a convolutional neural network (CNN) for the estimation of IVIM parameters. Approach: A CNN was trained on 338 images from the San Antonio Mexican American Family Study cohort and tested on 12 subjects at baseline and 12-week follow-up. Results: The CNN demonstrated improved repeatability for D* (wCV: 9.11% v. 19.3%) and D (wCV: 6.07% v. 10.7%) compared to the conventional non-linear least squares method. Impact: This study showed that CNNs improve the repeatability of D* and D estimates in the liver, though it remains unclear if the within-subject variability of IVIM parameters is sufficient to accurately differentiate fibrosis stages.