Machine Learning Enabled Non-invasive Diagnosis of Nonalcoholic Fatty Liver Disease and Assessment of Abdominal Fat from MRI Data
Pillai, A.; Bliznashki, K.; Hutchison, E.; Kumar, C.; Challis, B.; Patel, M.
Show abstract
Nonalcoholic fatty liver disease (NAFLD) is the most rapidly growing contributor to chronic liver disease worldwide with high disease burden and suffers from limitations in diagnosis. Inspired by recent advances in machine learning digital diagnostics, we explored the efficacy of training a neural network to classify high risk NAFLD vs. non-NAFLD patients in the UK Biobank dataset based on proton density fat fraction (PDFF). We compared the performance of several ResNet-derived architectures in the context of whole abdomen MRI, segmented liver and abdomen excluding liver (sans-liver). Non-local ResNet trained on whole abdomen MRI images yielded the highest precision (0.88 for NAFLD) and F1 (0.89 for NAFLD). Furthermore, our work on a second, larger cohort explored multi-task learning and the relationship among PDFF, visceral adipose tissue (VAT) and abdominal subcutaneous adipose tissue (ASAT). Interestingly, multi-task learning experiments found a decline in performance for PDFF when combined with VAT and ASAT. We address this deterioration using Multi-gate Mixture-of-Experts (MMoE) approaches. Our work opens the possibility for using a non-invasive deep learning-based diagnostic for NAFLD, and directly enables clinical and genomic research using a larger cohort of potential NAFLD patients in the UK Biobank study.
Matching journals
The top 9 journals account for 50% of the predicted probability mass.