Multiphase CT-based differentiation of cholangiocarcinoma using supervised contrastive learning

Abstract Background Cholangiocarcinoma (CCA) is a malignant biliary tumor with poor prognosis whose imaging findings on computed tomography (CT) frequently overlap with those of benign hepatobiliary conditions, making differential diagnosis clinically challenging, particularly for perihilar disease. Prior CT-based deep-learning studies of CCA have largely relied on single-phase images and expert-defined two-dimensional regions of interest, which limit the capture of three-dimensional biliary anatomy and restrict practical deployment. We aimed to develop and evaluate a fully automated system for differentiating CCA from benign hepatobiliary diseases on multiphase contrast-enhanced CT using supervised contrastive learning. Methods Automated liver and gallbladder segmentation was applied to extract three-dimensional hepatobiliary patches covering intrahepatic and extrahepatic bile ducts. Arterial- and delayed-phase patches from the same patient were used as positive pairs for supervised contrastive pre-training, followed by classifier fine-tuning on portal venous phase patches. The system was retrospectively evaluated in 427 patients (129 CCA, 179 benign gallbladder disease [BGD], 119 benign bile duct disease [BBD]) using stratified five-fold cross-validation. Diagnostic performance was assessed using sensitivity, specificity, balanced accuracy (bAcc), and area under the ROC curve (AUC), with results summarized across the five cross-validation folds. Results The model achieved a bAcc of 82.43% (AUC 89.79%) for differentiating CCA from BGD, and a bAcc of 69.90% (AUC 78.26%) for differentiation from BBD. In both tasks the proposed model achieved a significantly higher pooled AUC than all single-phase models (all Holm-adjusted $$p < 0.05$$ ) and consistently exceeded the multiphase early-fusion baseline across five repeated cross-validation runs, although this difference did not reach statistical significance in the primary analysis. Grad-CAM visualizations localized prediction-relevant features to periductal and intrahepatic parenchymal regions, supporting the anatomical plausibility of the model’s predictions. Conclusions Phase-aware contrastive learning with automated 3D patch extraction enhances CT-based differentiation of CCA from benign hepatobiliary conditions. This approach may serve as a complementary decision-support tool to improve diagnostic consistency, particularly for challenging cases such as perihilar cholangiocarcinoma, and shows potential for integration into clinical workflows.