Application of intratumoral and peritumoral radiomics based on multimodal MRI in predicting the pathological grade of intrahepatic mass-forming cholangiocarcinoma

Objective This work aimed to develop and internally validate a retrospective, single-center multimodal MRI radiomics model combining intratumoral and peritumoral features for noninvasive preoperative discrimination between high (moderately/poorly differentiated) and low-grade (well differentiated) intrahepatic mass-forming cholangiocarcinoma (IMCC). Methods Clinical and imaging data from 335 patients with histopathologically confirmed IMCC were retrospectively analyzed, and they were split into a training cohort (n = 234) and a validation cohort (n = 101). All enrolled subjects underwent a preoperative multimodal MRI examination, including T2WI, DWI, and DCE sequences. Manual contouring of ROIs was performed inside each tumor, whereas peritumoral ROIs were automatically generated at expansion distances of 5 mm, 10 mm, and 15 mm from the tumor boundary, followed by radiomic feature extraction. Dimensionality reduction and feature selection were performed using the LASSO regression method. Radiomic analysis extracted 6741 features per patient from multimodal MRI; after reproducibility filtering and LASSO selection, 15 high-value intratumoral/peritumoral features were retained to build radiomic signatures. Three Rad-scores were subsequently established: intratumoral, peritumoral, and the combined version (intratumoral + peritumoral). A multivariate logistic regression approach was applied to build a combined predictive model. The area under the receiver operating characteristic curve (AUC), calibration plots, and DCA were used as evaluation metrics for model performance. Results From the entire feature pool, fifteen optimal radiomic features were ultimately chosen, among which the DCE sequence was the greatest contributor (60.0%). Peritumoral features occupied 73.3% of all selected features, and the 10 mm peritumoral extension emerged as having the highest predictive capability. Regarding the identification of high pathological grade (i.e., moderately to poorly differentiated lesions), the combined radiomic signature delivered AUC values of 0.904 (95% CI: 0.862–0.946) in the training cohort and 0.879 (95% CI: 0.818–0.940) in the validation cohort, demonstrating a statistically superior performance relative to either the intratumoral or peritumoral models used alone ( P < 0.05). Following the incorporation of independent clinical risk indicators (namely, maximum tumor diameter and CA19-9 level), the resulting nomogram model yielded enhanced AUCs of 0.931 (95% CI: 0.893–0.969) within the training set and 0.907 (95% CI: 0.856–0.958) within the validation set. Satisfactory consistency was reflected by the calibration curves, while DCA indicated a favorable clinical net benefit. Conclusion The multimodal MRI-driven combined radiomic model, which drew from both intratumoral and peritumoral features, provided a powerful approach for preoperatively predicting the pathological grade of IMCC. The modeling strategy integrating DCE sequence features and the 10 mm peritumoral region demonstrated optimal performance, providing a noninvasive imaging tool for precise preoperative stratification and individualized treatment decisions. While performance was high in internal validation, the retrospective single-center design, single vendor MRI, and lack of external validation limit immediate generalizability, and multicenter prospective studies are needed.