Comparative multiphase CT radiomics for differentiating isolated hepatic metastasis from intrahepatic cholangiocarcinoma: a two-center study of native, subtraction, peritumoral, and fusion models.

Background: Isolated hepatic metastasis (IHM) and intrahepatic cholangiocarcinoma (IMCC) may show overlapping appearances on multiphasic contrast-enhanced computed tomography (CT). A quantitative imaging approach may help improve noninvasive differentiation. This study aims to develop and externally validate multiphasic contrast-enhanced CT radiomics models for differentiating IHM from IMCC and to evaluate the diagnostic value of native, subtraction, peritumoral, and fusion information. Methods: This retrospective two-center study included patients with pathologically confirmed IHM or IMCC who underwent multiphasic liver CT at Huzhou Central Hospital or Yijishan Hospital Affiliated to Wannan Medical University between January 2023 and December 2025. The final cohort included 227 patients. Center 1 was used as the training cohort (n=115; IHM, n=68; IMCC, n=47), and center 2 was used as the external validation cohort (n=112; IHM, n=60; IMCC, n=52). Clinical baseline variables included sex, age, and serum tumor markers. Radiomics features were extracted using PyRadiomics from unenhanced CT images (C), arterial phase CT images (A), venous phase CT images (V), and subtraction images M=A-C, N=V-C, and P=A-V. Ten prespecified models were constructed to represent native-image, subtraction-image, peritumoral, and fusion information levels. Feature selection was performed in the training cohort using median imputation, variance filtering, and ANOVA F-test-based SelectKBest. XGBoost classifiers were tuned using five-fold stratified cross-validation in the training cohort and evaluated once in the external validation cohort. Results: In the training cohort, apparent AUCs ranged from 0.864 to 0.978. The subtraction-total model based on M+N+P intratumoral features showed the highest numerical external validation performance, with an AUC of 0.835 (0.755-0.907), accuracy of 0.759, sensitivity of 0.800, and specificity of 0.712. The intra-plus-peritumoral model achieved an external validation AUC of 0.800 (0.713-0.882). The plain unenhanced model achieved an AUC of 0.821 (0.739-0.900), whereas the native C+A+V model and ultimate fusion model achieved AUCs of 0.809 (0.725-0.888) and 0.800 (0.714-0.884), respectively. Pairwise DeLong tests among XGBoost models did not show statistically significant differences after multiple-comparison correction. Conclusion: Multiphase contrast-enhanced CT radiomics can differentiate IHM from IMCC. Subtraction-based intratumoral models demonstrated the highest numerical external validation performance, although the advantage was not statistically significant after multiple-comparison correction. These findings indicate the potential discriminatory value of enhancement-difference information.