Machine learning modeling to predict HCC locations in cirrhotic patients undergoing MRI - a proof-of-concept study.
Olivia Gaddum1,2, Tal Zeevi1, Weicheng Dai1
1Department of Radiology and Biomedical Imaging, Yale University, New Haven, USA.
Abdominal Radiology (New York)
|July 13, 2026
Summary
Machine learning (ML) models can identify liver regions at high risk for hepatocellular carcinoma (HCC) using MRI scans. This approach aids in early detection before radiologic diagnosis, especially when precursor lesions are present.
Area of Science:
- Radiology and Medical Imaging
- Machine Learning in Medicine
- Hepatocellular Carcinoma (HCC) Research
Background:
- Current hepatocellular carcinoma (HCC) surveillance relies on subjective LI-RADS features, not direct imaging analysis.
- There's a need for objective, data-driven methods to identify at-risk liver parenchyma for HCC.
- Machine learning (ML) offers potential for automated analysis of medical imaging data.
Purpose of the Study:
- To evaluate the feasibility of ML-based image analysis frameworks for identifying and localizing hepatic parenchyma at elevated risk for HCC.
- To assess ML model performance in predicting HCC development using MRI data.
- To explore the utility of ML in improving HCC surveillance beyond current manual methods.
Main Methods:
- Retrospective analysis of cirrhotic patients with HCC diagnosis undergoing MRI (2008-2023).
- Inclusion of screening MRIs preceding a confirmed LR-5 lesion within 18 months.
- Radiomic feature extraction from manually/automatically annotated 'non-malignant' and 'malignant' liver tissue VOIs; training and validation of Logistic Regression (LR), Random Forest (RF), and eXtreme Gradient Boosting (XGB) models.
Main Results:
- Best model performances (AUC) for LR, RF, and XGB were 0.75, 0.80, and 0.79, respectively, using manual annotations.
- RF model achieved an AUC of 0.86 for LR-3 lesions, outperforming regions without precursor lesions (AUC 0.66).
- Voxel-level heatmaps indicated increased HCC probability in subsequent lesion locations in 30% of patients.
Conclusions:
- ML-based MRI analysis is feasible for identifying liver regions with increased HCC risk before radiologic diagnosis.
- The models show particular promise in the presence of precursor lesions.
- This approach could enhance early HCC detection and surveillance strategies.
