Related Experiment Video For atypical focal hepatic lesion
Updated: May 14, 2026
Predicting Treatment Response to Image-Guided Therapies Using Machine Learning: An Example for Trans-Arterial Treatment of Hepatocellular Carcinoma
Published on: October 10, 2018
A Clinlabomics-Based Machine Learning Model Accurately Differentiates Atypical Hepatocellular Carcinoma from Atypical
Qing-Qing Luo1, Ding-Fan Guo1, Qiao-Nan Li1
1Department of Gastroenterology, Jiangxi Provincial Key Laboratory of Digestive Diseases, Jiangxi Clinical Research Center for Gastroenterology, Digestive Disease Hospital, The First Affiliated Hospital, Jiangxi Medical College, Nanchang University, Nanchang, Jiangxi, People's Republic of China.
Purpose:
The differential diagnosis of radiological image atypical hepatocellular carcinoma (aHCC) and atypical benign focal hepatic lesion (aBFHL) is a challenge. We aimed to develop a diagnostic model based on the new concept of clinlabomics to address this challenge.
Patients And Methods:
Pathologically diagnosed 466 patients (252 aHCC and 214 aBFHL) and their clinlabomic data were retrospectively collected. The patients were split into two sets based on admission time for training and testing models to differentiate aHCC from aBFHL. The models were developed using the three best-performing algorithms and key features selected from 18 clinlabomic indicators. The best model was validated and evaluated for classification ability by receiver operating characteristic (ROC) curve, fitness by calibration curve, and clinical utility by decision curve. Model interpretability was analyzed through SHapley Additive exPlanations, and model application was realized via an easy-to-use online calculator.
Results:
Random forest (RF), support vector machine, and linear discriminant analysis were the top three algorithms. Six features (hepatitis B surface antigen, sex, alpha-fetoprotein, aspartate aminotransferase, platelets, and age) were selected. Three models were developed using the selected algorithms and features, and the RF model was the optimal one, with an area under the ROC curve (AUC) of 0.954 and a diagnostic accuracy of 92.5% for the testing set. Notably, this model also outperformed for early-stage, small, and AFP-negative aHCCs, with AUCs of 0.976-0.982 and accuracies of 93.2-93.8%. The RF model performed well in terms of calibration, net benefit gain, interpretability, and application.
Conclusion:
The clinlabomics-based diagnostic model is valuable in the differential diagnosis of various types of aHCC from aBFHL, including early-stage, small, and AFP-negative aHCCs.
