Deep Learning Reveals Liver MRI Features Associated With PNPLA3 I148M in Steatotic Liver Disease

Yazhou Chen, Benjamin P. M. Laevens, Teresa Lemainque, Gustav Müller‐Franzes, Tobias Seibel, Carola Dlugosch, Jan Clusmann, Paul‐Henry Koop, Rongpeng Gong, Yuanyuan Liu, Niharika Jakhar, Feng Cao, Simon Schophaus, Thriveni Basavanapura Raju, Anastasia Raptis, Felix van Haag, Joel Joy, Rohit Loomba, Luca Valenti, Jakob Nikolas Kather, Titus J. Brinker, Moritz Herzog, Ivan G. Costa, Diego Hernando, Kai Markus Schneider, Daniel Truhn, Carolin V. SchneiderLiver 2025 · DOI

TL;DR

This paper introduces a deep learning model that uses liver MRI scans to predict whether a person carries a specific genetic variant (PNPLA3 I148M) associated with steatotic liver disease (SLD). This non-invasive approach could pave the way for earlier diagnosis and personalized treatment strategies for individuals at higher risk of developing severe liver conditions.

What Problem Does This Paper Address?

Steatotic liver disease (SLD), particularly metabolic dysfunction-associated steatotic liver disease (MASLD), is a growing global health concern affecting a significant portion of the population. Genetic factors play a crucial role in its development and progression. The PNPLA3 I148M variant is a key genetic contributor, increasing the risk of SLD two to three-fold. However, genetic testing for this variant is not standard clinical practice, preventing personalized care. This study aims to bridge this gap by developing a non-invasive method to identify carriers of the PNPLA3 I148M variant using readily available liver MRI data, which is often acquired for other clinical reasons.

Methods

The researchers employed a deep learning approach using a Vision Transformer (ViT) model. Here's a breakdown of their key methods:

  • Data Preprocessing: Liver MRI data was first processed to generate proton density fat fraction (PDFF) maps using a water-fat separation toolbox. A U-Net model was used to segment the liver region from these images.
  • Cohort Selection: The study focused on individuals from the UK Biobank. To ensure a clear genetic contrast, only homozygous carriers and non-carriers of the PNPLA3 I148M variant were included, excluding heterozygotes. Individuals were then categorized into SLD (PDFF ≥ 5%) and non-SLD (PDFF < 5%) groups.
  • Model Architecture: A Vision Transformer (ViT-base-patch16-224) model was chosen for its effectiveness in image classification tasks.
  • Training and Validation: The model was trained using a five-fold stratified cross-validation approach. An internal test set (15% of the data) was used for evaluation.
  • Interpretation: Attention maps were generated to visualize the regions within the liver MRI that were most predictive of the PNPLA3 I148M variant status, offering insights into why the model made its predictions.

Datasets

The study utilized data from the UK Biobank, a large-scale biomedical database containing comprehensive genetic and health information from over half a million UK participants. Specifically, they used liver MRI data from a subset of 45,603 individuals acquired using a multi-echo spoiled gradient echo sequence.

Key Insights and Ideas

  • MRI as a Non-Invasive Genetic Predictor: The core idea is that subtle patterns in liver fat distribution, visible on MRI, can be indicative of underlying genetic predispositions like the PNPLA3 I148M variant.
  • Vision Transformers for Medical Imaging: The successful application of Vision Transformers, a powerful deep learning architecture, to medical imaging for genetic variant prediction demonstrates their potential beyond traditional computer vision tasks.
  • Location Matters: The study found that the location of liver fat, particularly in areas adjacent to blood vessels in the liver's central region, was more predictive than just the overall amount of fat. This suggests a more nuanced biological mechanism at play.
  • PNPLA3 I148M's Unique Signature: The PNPLA3 I148M variant showed a more distinct and predictable pattern on MRI compared to other steatosis-associated genetic variants examined, highlighting its significant impact on liver fat distribution.

Key Results

  • Predictive Performance: The Vision Transformer model achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.68 for predicting PNPLA3 I148M homozygosity in the SLD group. While lower in the full cohort (0.61) and non-SLD group (0.57), this still represents a significant ability to distinguish carriers.
  • Variant Specificity: The PNPLA3 I148M variant consistently showed the highest predictive performance among the five genetic variants tested.
  • Attention Map Insights: Attention maps highlighted specific regions of fat deposition, particularly near hepatic vessels and the liver hilum, as crucial for prediction.
  • Impact of PDFF and ALT: Higher PDFF (liver fat content) and ALT (liver enzyme) levels were associated with better prediction accuracy, reinforcing the link between these markers and the PNPLA3 I148M variant.
  • Model Robustness: The model's predictions were not significantly driven by general risk factors like obesity or Type 2 diabetes, suggesting it captures genotype-specific imaging signatures.

Limitations and Future Work

  • Generalizability: The study acknowledges that conclusions may not be directly applicable to non-European populations or individuals with advanced liver fibrosis, where steatosis might be less pronounced.
  • External Validation: The lack of an external cohort for validation is a significant limitation. Future work should include external validation to confirm the model's generalizability.
  • Image Quality: The quality of MRI images can influence prediction outcomes. Future research could benefit from higher-resolution 2D or 3D liver data.
  • Integration with Other Data: The authors suggest that combining MRI-based predictions with clinical and biochemical data (like BMI, serum values) could further enhance accuracy and clinical utility.
  • Radiomics: Integrating radiomics features extracted from liver regions alongside imaging data could provide a more comprehensive set of predictive features.

Summary updated

Abstract

Our study marks novel progress in the non-invasive detection of homozygosity for PNPLA3 I148M through the application of deep learning models on MRI images. Our findings suggest that PNPLA3 I148M might affect the liver fat distribution and could be used to predict the presence of PNPLA3 variants in patients with fatty liver. The findings of this research have the potential to be integrated into standard clinical practice, particularly when combined with clinical and biochemical data from other modalities to increase accuracy, enabling easier identification of at-risk individuals and facilitating the development of tailored interventions for PNPLA3 I148M-associated liver disease.
BibTeX
@inproceedings{chen2025deep,
  author = {Yazhou Chen and Benjamin P. M. Laevens and Teresa Lemainque and Gustav Müller‐Franzes and Tobias Seibel and Carola Dlugosch and Jan Clusmann and Paul‐Henry Koop and Rongpeng Gong and Yuanyuan Liu and Niharika Jakhar and Feng Cao and Simon Schophaus and Thriveni Basavanapura Raju and Anastasia Raptis and Felix van Haag and Joel Joy and Rohit Loomba and Luca Valenti and Jakob Nikolas Kather and Titus J. Brinker and Moritz Herzog and Ivan G. Costa and Diego Hernando and Kai Markus Schneider and Daniel Truhn and Carolin V. Schneider},
  title = {{Deep Learning Reveals Liver MRI Features Associated With <i>PNPLA3</i> I148M in Steatotic Liver Disease}},
  booktitle = {Liver},
  year = {2025},
  doi = {10.1111/liv.70164},
}
Figure