Machine learning–based multiphase CT radiomics for differentiating hepatocellular carcinoma from hepatocellular adenoma in dogs

Noninvasive differentiation between hepatocellular carcinoma (HCC) and hepatocellular adenoma (HCA) is clinically important. Although computed tomography (CT) is widely used to evaluate hepatic masses, these entities often exhibit overlapping imaging features, limiting differentiation based on conventional CT findings alone. Given this limitation, we evaluated conventional CT features and developed radiomics-based machine learning (ML) models to differentiate HCC from HCA in dogs. Seventy-five dogs with histologically confirmed HCC or HCA that underwent triple-phase CT were retrospectively included. For conventional CT evaluation, regions of interest (ROIs) were manually segmented for each hepatic mass across all phases. In addition, 1,316 radiomic features were extracted from the segmented ROIs. SHAP (Shapley Additive Explanations)-based feature selection was applied, and the selected features were evaluated using multiple ML classifiers. Conventional CT analysis demonstrated limited qualitative differences between HCC and HCA. In contrast, radiomics-based ML analysis showed high cross-validated classification performance, particularly with random forest and boosted tree-based models. Washout-related intensity and texture features in the delayed phase were the most informative, and multiphase feature integration achieved the highest performance. In summary, radiomics-based ML analysis of triple-phase CT showed promising potential for differentiating HCC from HCA, suggesting that it may serve as a noninvasive adjunct for the preoperative diagnosis of canine hepatocellular tumors.