Related Experiment Video
Updated: Aug 22, 2025
08:41
Novel In Vivo Micro-Computed Tomography Imaging Techniques for Assessing the Progression of Non-Alcoholic Fatty Liver Disease
Published on: March 24, 2023
1.3K
Deep learning-based quantification of NAFLD/NASH progression in human liver biopsies
Fabian Heinemann1, Peter Gross2, Svetlana Zeveleva3
1Drug Discovery Sciences, Boehringer Ingelheim Pharma GmbH & Co. KG, 88397, Biberach an der Riß, Germany. fabian.heinemann@boehringer-ingelheim.com.
Scientific Reports
|November 10, 2022
Summary
A new deep learning method automates the analysis of liver biopsies for non-alcoholic fatty liver disease (NAFLD) and non-alcoholic steatohepatitis (NASH). This AI approach offers reproducible, high-resolution scoring, improving upon traditional pathologist assessments.
Area of Science:
- Hepatology
- Medical Imaging
- Artificial Intelligence
Background:
- Non-alcoholic fatty liver disease (NAFLD) impacts 24% globally, potentially progressing to NASH, cirrhosis, and liver cancer.
- Current diagnosis relies on liver biopsy and pathologist analysis, which is time-consuming and subject to variability.
- The Kleiner score quantifies steatosis, inflammation, ballooning, and fibrosis, but has limitations in resolution and reproducibility.
Purpose of the Study:
- To develop and validate an automated deep learning method for analyzing liver biopsies.
- To provide reproducible and higher-resolution scoring of histopathological features in NAFLD/NASH.
- To create a continuous scoring system comparable to pathologist-derived scores.
Main Methods:
- Developed a deep learning system trained on 296 human liver biopsies.
- Tested the system on 171 human liver biopsies with pathologist ground truth scores.
- The AI mimics pathologist analysis by first identifying features and then aggregating them into a per-biopsy score.
Main Results:
- The automated system achieved scores comparable to pathologist ground truth.
- Quadratic weighted Cohen's κ values ranged from 0.24 (inflammation) to 0.66 (steatosis).
- Mean absolute errors were lowest for steatosis (0.29) and highest for NAS (0.77).
Conclusions:
- The deep learning method offers a reproducible and high-resolution alternative for assessing NAFLD/NASH.
- This automated approach can assist pathologists in liver biopsy analysis.
- The system's continuous scoring scale provides a more granular assessment of disease progression.
