Multi-regional radiomics for predicting microvascular invasion and lymph node metastasis in intrahepatic cholangiocarcinoma.

AIM: To evaluate the predictive value of multiregional radiomics signatures for preoperative microvascular invasion (MVI) status and lymph node metastasis (LNM) in patients with intrahepatic cholangiocarcinoma (ICC).

MATERIAL AND

METHODS

This study included 200 ICC patients (training cohort: n = 160; validation cohort: n = 40) who underwent preoperative contrast-enhanced magnetic resonance imaging (MRI). For each patient, volumes of interest (VOIs) were defined for the intratumoural region (VOItumor) and combined intratumoural with peritumoural 8, 10, and 12 mm regions (VOI8mm, VOI10mm, VOI12mm). The least absolute shrinkage and selection operator (LASSO) method was applied to screen radiomics features. Optimal radiomics signatures and independent risk predictors were incorporated into the combined models. To address the imbalance in LNM data, synthetic minority oversampling technique (SMOTE) and random under sampler (RUS) were used for subsampling and modelled with logistic regression and k-nearest neighbour (KNN) classifiers.

RESULTS

The MVI combined model incorporated radiomics score (Radscore) of VOI10mm, intrahepatic duct dilatation, and tumour size. It achieved satisfactory prediction performance in the validation cohort: area under the curve (AUC) = 0.827. For LNM prediction, the model (SMOTE + KNN), which combined carcinoembryonic antigen (CEA) levels with Radscoretumor performed best in the validation cohort: area under the receiver operating characteristic curve (AUROC) = 0.853, area under the precision-recall curve (AUPRC) = 0.749.

CONCLUSION

The Radscore10mm showed effective MVI discrimination performance, and the final combined model showed superior discrimination. With Radscoretumor and CEA, the tumour model can effectively predict LNM, and SMOTE enhances the performance of the model.