Preoperative differentiation of primary liver cancer subtypes: An MRI-based multiclass classification model integrating radiomic and clinicoradiological characteristics

This study developed an magnetic resonance imaging (MRI)-based multiclass model for the preoperative differentiation of primary liver cancer (PLC) subtypes by integrating radiomic and clinicoradiological features. We retrospectively enrolled 251 patients with pathologically confirmed PLC, including 92 with hepatocellular carcinoma, 77 with intrahepatic cholangiocarcinoma, and 82 with combined hepatocellular-cholangiocarcinoma. A clinicoradiological model was constructed using statistically significant clinical and MRI-based radiological characteristics. To develop the radiomics model, radiomic features were extracted from multiparametric MRI sequences, including diffusion-weighted imaging, dynamic contrast-enhanced T1WI in the arterial (T1WI-A), portal venous (T1WI-V), and delayed (T1WI-D) phases. A combined model was then developed by integrating these features. Feature selection employed a 2-stage strategy: initial filtering coupled with model construction using 6 machine learning (ML) algorithms, followed by recursive feature elimination (RFE) on the top-performing model to build the final RFE-integrated model. Performance was assessed by the area under the receiver operating characteristic curve (AUC), sensitivity, accuracy, precision, and specificity. Additionally, SHapley Additive exPlanations values were applied to interpret the model's prediction logic. The RFE-integrated model based on a support vector machine demonstrated the highest overall and per-class classification performance, achieving micro- and macro-average AUCs of 0.934 (95% CI: 0.908-0.955) and 0.925 (95% CI: 0.896-0.949), along with a sensitivity of 0.805 (95% CI: 0.761-0.850) and a specificity of 0.901 (95% CI: 0.878-0.925). The model also showed high discriminatory power for individual subtypes with AUCs of 0.931 for hepatocellular carcinoma, 0.975 for intrahepatic cholangiocarcinoma, and 0.868 for combined hepatocellular-cholangiocarcinoma. In conclusion, these findings demonstrate that the support vector machine-based RFE-integrated model provides a highly accurate, noninvasive tool for preoperative PLC subtyping, addressing a critical diagnostic challenge, facilitating subtype-specific management, and potentially improving patient outcomes.