Application of A U-Net for Map-like Segmentation and Classification of Discontinuous Fibrosis Distribution in Gd-EOB-DTPA-Enhanced Liver MRI

Abstract

We aimed to evaluate whether U-shaped convolutional neuronal networks can be used to segment liver parenchyma and indicate the degree of liver fibrosis/cirrhosis at the voxel level using contrast-enhanced magnetic resonance imaging. This retrospective study included 112 examinations with histologically determined liver fibrosis/cirrhosis grade (Ishak score) as the ground truth. The T1-weighted volume-interpolated breath-hold examination sequences of native, arterial, late arterial, portal venous, and hepatobiliary phases were semi-automatically segmented and co-registered. The segmentations were assigned the corresponding Ishak score. In a nested cross-validation procedure, five models of a convolutional neural network with U-Net architecture (nnU-Net) were trained, with the dataset being divided into stratified training/validation (n = 89/90) and holdout test datasets (n = 23/22). The trained models precisely segmented the test data (mean dice similarity coefficient = 0.938) and assigned separate fibrosis scores to each voxel, allowing localization-dependent determination of the degree of fibrosis. The per voxel results were evaluated by the histologically determined fibrosis score. The micro-average area under the receiver operating characteristic curve of this seven-class classification problem (Ishak score 0 to 6) was 0.752 for the test data. The top-three-accuracy-score was 0.750. We conclude that determining fibrosis grade or cirrhosis based on multiphase Gd-EOB-DTPA-enhanced liver MRI seems feasible using a 2D U-Net. Prospective studies with localized biopsies are needed to evaluate the reliability of this model in a clinical setting.