Deep Learning for Liver Disease Stratification: Findings from UKBB MRI

Rami Al-Belmpeisi, Josefine Vilsbøll Sundgaard, Peter Hjørringgaard Larsen, Anders Dahl

Keywords: Deep Learning, MRI, Explainability, Learning with Noisy Labels, Classification of medical images, Segmentation of medical images

TL;DR: Deep Learning method for diagnosing MASLD and MASH from UK Biobank MRI data with noisy labels.

Abstract:

Metabolic dysfunction-associated steatotic liver disease (MASLD) and its progressive form, metabolic dysfunction-associated steatohepatitis (MASH), have become more prevalent, spurring interest in using magnetic resonance imaging (MRI) sequences for diagnosis.
In this study, we propose a method that uses deep learning to diagnose MASLD and MASH with significant fibrosis from single-slice (2D) and volumetric (3D) MRI sequences that originate from the UK Biobank. In this paper, we focus on transparent decision-making.
Our study shows that imposing anatomically informed constraints by using a liver segmentation mask on the network's input has minimal impact on diagnostic performance. Still, it redirects attention to clinically relevant liver regions, preventing shortcut learning from extrahepatic features, such as subcutaneous fat.
These constraints shift the focus of the model toward proton density fat fraction (PDFF) maps for healthy liver assessment, $T_1$ maps for MASLD diagnosis, and both sequences to identify MASH with significant fibrosis.
Our top-performing models achieve AUCs of 0.89/0.96/0.79 for the diagnosis of the healthy/MASLD/MASH groups with significant fibrosis, respectively.
Despite label noise and limited sequence specificity, which primarily hinder predictive performance in cases of MASH with significant fibrosis, the identified indicators are frequently located in liver regions consistent with prior understanding of disease progression.
In conclusion, we find that 2D MRI sequences are sufficient for diagnosing MASLD/MASH with significant fibrosis, as performance decreases and computation time increases when using 3D volumes.

Primary Subject Area: Learning with Noisy Labels and Limited Data

Secondary Subject Area: Interpretability and Explainable AI

Registration Requirement: Yes

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Midl Latex Submission Checklist: Ensure no LaTeX errors during compilation., Created a single midl26_NNN.zip file with midl26_NNN.tex, midl26_NNN.bib, all necessary figures and files., Includes \documentclass{midl}, \jmlryear{2026}, \jmlrworkshop, \jmlrvolume, \editors, and correct \bibliography command., Did not override options of the hyperref package, Did not use the times package., All authors and co-authors are correctly listed with proper spelling and avoid Unicode characters., Author and institution details are de-anonymized where needed. All author names, affiliations, and paper title are correctly spelled and capitalized in the biography section., References must use the .bib file. Did not override the bibliographystyle defined in midl.cls. Did not use \begin{thebibliography} directly to insert references., Tables and figures do not overflow margins; avoid using \scalebox; used \resizebox when needed., Included all necessary figures and removed *unused* files in the zip archive., Removed special formatting, visual annotations, and highlights used during rebuttal., All special characters in the paper and .bib file use LaTeX commands (e.g., \'e for é)., Appendices and supplementary material are included in the same PDF after references., Validation paper does not exceed 16 pages; acknowledgements, references, and appendix start on page 17 or later.

Submission Number: 28