Abstract / Summary
Abstract Background Esophagogastric variceal bleeding (EGVB) is a life-threatening complication of cirrhosis. Non-contrast computed tomography (NCCT) is free of contrast-related risks, but it cannot directly visualize varices. We therefore developed and validated a fusion model that integrates radiomics with a three-dimensional convolutional neural network (3D CNN) to predict EGVB risk from NCCT. Methods A total of 137 cirrhotic patients (47 with prior EGVB, 90 without) who underwent NCCT were retrospectively enrolled. Radiomics features (n = 1,919) were extracted from seven ROIs (esophagus, gastric fundus, liver, spleen, portal vein, splenic vein, superior mesenteric vein). For the fusion model, the deep features were concatenated with the LASSO-selected radiomics features from the same training fold. A 3D ResNet-18 CNN learned 128 deep features from the same NCCT volumes. The two feature sets were concatenated and classified by a multilayer perceptron to form the fusion model. Five-fold cross-validation was used for performance evaluation, including AUC, calibration and decision curve analysis. Results The fusion model achieved a mean AUC of 0.884 ± 0.037 across five folds, with a sensitivity of 0.668 and specificity of 0.944 at the Youden-index optimal threshold. The calibration curve showed close agreement between predicted and observed probabilities. Decision curve analysis demonstrated consistent net benefit over both treat-all and treat-none strategies across the threshold range of 0.10–0.50, suggesting potential clinical utility. Conclusions The NCCT-based fusion model integrating radiomics and 3D CNN provides good discrimination and calibration for EGVB risk stratification, and may serve as a non-invasive tool to guide prophylactic decisions. Larger prospective multi-center studies are warranted to confirm its clinical utility.