Machine learning modeling to predict HCC locations in cirrhotic patients undergoing MRI - a proof-of-concept study.
Authors
Affiliations (11)
Affiliations (11)
- Department of Radiology and Biomedical Imaging, Yale University, New Haven, USA.
- Department of Radiology, Charité Universitätsmedizin Berlin, Corporate Member of Freie Universität Berlin and Humboldt-Universität, Berlin, Germany.
- Visage Imaging Inc., California, San Diego, USA.
- Department of Radiology, University of California, San Diego, USA.
- Rudolf Virchow Center for Integrative and Translational Biomedicine, Julius Maximilians University of Wuerzburg, Wuerzburg, Germany.
- Institute of Experimental Biomedicine, University Hospital Wuerzburg, Wuerzburg, Germany.
- Experimental Clinical Research Center (ECRC), Charité-Universitätsmedizin Berlin and Max-Delbrück-Centrum für Molekulare Medizin (MDC), Berlin, Germany.
- Berlin Institute of Health, Charité-Universitätsmedizin Berlin, Berlin, Germany.
- Department of Radiology and Nuclear Medicine, Otto-von-Guericke University Magdeburg, Magdeburg, Germany.
- Biomedical Engineering, Radiology & Biomedical Imaging, Electrical Engineering and Statistics & Data Science, Yale University, Connecticut, New Haven, USA.
- Department of Radiology and Biomedical Imaging, Yale University, New Haven, USA. [email protected].
Abstract
Current hepatocellular carcinoma (HCC) surveillance guidelines rely on manually defined LI-RADS (Liver Imaging Reporting and Data System) features rather than imaging data analysis. This study evaluates the feasibility of machine-learning (ML)-based image analysis frameworks to identify and localize hepatic parenchyma at elevated risk for HCC. In this retrospective study, cirrhotic patients with HCC diagnosis undergoing MRI between 2008 and 2023 were included. The analysis included negative screening MRI preceding a positive screening MRI confirming a LR-5 lesion within 18 months. Volume-of-interest (VOI) annotations of 'non-malignant' and 'malignant' liver tissue on the screening MRI were manually or automatically placed, mapped from future HCC lesion on positive screening MRI. Radiomics were extracted from these VOIs using PyRadiomics. Logistic regression (LR), random forest (RF), and eXtreme Gradient Boosting (XGB) models were trained and validated across four manual/automatic annotation combinations. Exploratory voxel-level heatmaps were generated to visualize high-risk HCC areas. Model performance was summarized using median values and 95% non-parametric confidence intervals. 121 patients (65 ± 9.56; 99 men) were included. The best model performances for LR, RF, and XGB were achieved when training and validating on manual annotations: (AUC [95% CI]: 0.75 [0.72-0.78]; 0.80 [0.78-0.81]; 0.79 [0.77-0.81], respectively). Predictions from LR-3 lesions outperformed regions without visible precursor territory (RF AUC [95% CI]: 0.86 [0.84-0.88] vs. 0.66 [0.63-0.70]). In 36 out of 121 patients (30%), heatmaps showed visually increased probability signals in the region where HCC subsequently became visible on follow-up imaging. The results demonstrate the feasibility of ML-based MRI analysis for identifying liver regions at increased risk for HCC development prior to radiologic diagnosis, particularly in the presence of precursor lesions.