Machine learning-based CT radiomics for identifying high risk esophageal varices of bleeding in hepatocellular carcinoma.
Authors
Affiliations (6)
Affiliations (6)
- Department of Medical Imaging and Intervention, Chang Gung Memorial Hospital, Linkou, No. 5, Fuxing St., Guishan Dist., 33305, Taoyuan City, Taiwan.
- Graduate Institute of Artificial Intelligence, College of Intelligent Computing, Chang Gung University, No. 259, Wenhua 1st Rd., Guishan Dist., 33302, Taoyuan City, Taiwan.
- Department of Artificial Intelligence, Chang Gung University, No. 259, Wenhua 1st Rd., Guishan Dist., 33302, Taoyuan City, Taiwan.
- Institute of Health Data Science, Chang Gung University, No. 259, Wenhua 1st Rd., Guishan Dist., 33302, Taoyuan City, Taiwan.
- Department of Chinese Acupuncture and Traumatology, Center of Traditional Chinese Medicine, Chang Gung Memorial Hospital, Taoyuan, No. 123, Dinghu Rd., Guishan Dist., 33378, Taoyuan City, Taiwan.
- Department of Medical Imaging and Intervention, Chang Gung Memorial Hospital, Linkou, No. 5, Fuxing St., Guishan Dist., 33305, Taoyuan City, Taiwan. [email protected].
Abstract
Esophageal variceal (EV) bleeding is a lethal complication in patients with hepatocellular carcinoma (HCC). Current risk stratification relies on invasive endoscopy or systemic surrogates of liver cirrhosis. This study aimed to develop a machine learning model utilizing EV-specific CT radiomics to identify high-risk esophageal varices, as defined by endoscopic criteria, in patients with HCC. 151 HCC patients were enrolled and divided into a training cohort (n = 101; 75 high-risk; January 2017 - December 2022) and an independent temporal validation cohort (n = 50; 29 high-risk; January 2023 - December 2024). Endoscopic stigmata were used as an endpoint, including active bleeding or varices with high-risk features. Radiomic features were extracted from contrast-enhanced CT images acquired within six months prior to endoscopy. Model performance was evaluated using the area under the curve (AUC) of receiver operating characteristic (ROC), Brier score, calibration curves, and decision curve analysis (DCA). In the training cohort, the varix-specific radiomics signature developed via the least absolute shrinkage and selection operator (LASSO) demonstrated excellent discriminative power (cross-validated AUC: 0.956; 95% CI 0.872-1.000). In the temporal validation cohort, this model maintained high generalizability, yielding an AUC of 0.926 (95% CI 0.846-0.984, sensitivity: 93.1%, specificity: 71.4%) as well as a Brier score of 0.1186 and good calibration (intercept = -0.157, slope = 1.354). Surface-to-Volume Ratio (SVR) and NGTDM Coarseness are feasible radiomics. The varix-specific radiomics model provides highly sensitive and mathematically stable non-invasive risk stratification for EV bleeding in patients with HCC.