Identifying Nonalcoholic Fatty Liver Disease and Advanced Liver Fibrosis from MRI in UK Biobank

Rami Al-Belmpeisi, Kristine Aavild Sørensen, Josefine Vilsbøll Sundgaard, Puria Nabilou, Monica Jane Emerson, Peter Hjørringgaard Larsen, Lise Lotte Gluud, Thomas Lund Andersen, Anders Bjorholm Dahl

Abstract: Non-alcoholic fatty liver disease (NAFLD) and its progressive form of non-alcoholic steatohepatitis (NASH) pose a major public health problem that affects more than 30% of the global population. Since NAFLD is asymptomatic in the early stages, sufferers often remain untreated until the onset of NASH, which can lead to fibrosis and eventually cirrhosis of the liver. This condition is traditionally diagnosed via liver biopsy, which is invasive and associated with significant risks for the patient and susceptibility to sampling errors. These limitations underscore the necessity for non-invasive tools to assess disease severity. We explore the potential of magnetic resonance imaging (MRI) sequences in the UK Biobank (UKBB) to classify individuals as having either a healthy liver, NAFLD, or progressive NAFLD-associated advanced fibrosis. For the classification inputs, we utilize proton density fat fraction (PDFF) and native spin-lattice relaxation time (T1) maps, as well as serum biomarker data for assessing the sub-cohorts. The best models achieve near-perfect performance on identifying healthy individuals and NAFLD with AUCs of 0.99 and 0.98 respectively, while individuals with advanced fibrosis are under-diagnosed with an AUC of 0.67 at best. While segmentation decreases model performance, when classifying on full images, we make use of non-liver-related features, which is sub-optimal if we want to detect liver-related imaging biomarkers.

External IDs:dblp:conf/mlmi-ws/AlBelmpeisiSSNELGAD24