Interpretable stacking model integrating intra-/peritumoral CT-radiomics and serum biomarkers for predicting microvascular invasion in HCC: a dual-center retrospective study.

purposeTo establish an interpretable stacking model integrating intratumoral and peritumoral multi-phase CT radiomics with serum biomarkers for predicting microvascular invasion (MVI) in hepatocellular carcinoma (HCC).

methodsWe retrospectively analyzed 219 HCC patients. 151 HCC patients from Center 1 were randomly divided into a training set (N = 105) and a test set (N = 46). 68 HCC patients from Center 2 were classified as an external validation set. Radiomics features were extracted from primary lesion and its surrounding 5 mm region on enhanced CT, and the total Stacking radiomics score (TOTALSRads) was constructed. The nutritional immune inflammation score (NIIs) was derived from serum biomarkers. Four machine learning classifiers were employed to develop models, and the optimal model (SLMXG) was integrated using several different fusion strategies. The model’s evaluation used the area under the curve (AUC), confidence interval (CI), and decision curve analysis (DCA). Model interpretability was assessed using SHAP analysis and correlation analyses.

resultsSLMXG achieved the best AUC of 0.9577 in the training set, and was well verified in test set (AUC: 0.8611[95%CI 0.7493–0.9729]) and external validation set (AUC: 0.8629[95%CI 0.7730–0.9528]) with good calibration. In external validation set, SLMXG exhibited an accuracy of 0.8088, precision of 0.8529, F1 score of 0.8169, and Brier score of 0.1545, showing favorable clinical net benefit with a threshold probability range of 20%‒85%. Correlation analyses indicated good agreement among intratumoral and peritumoral radiomics, TOTALSRads, and NIIs (P < 0.001). SHAP analyses highlighted that peritumoral radiomics and serum biomarkers contributed powerful enhancing roles for SLMXG, with Venous_phase_intra-peritumoral 0.5 cm radiomics and Xgboost contributing the most.

conclusionsThe SLMXG, which integrates different learning fusion strategies with serum biomarkers and intratumoral and peritumoral multi-phase CT radiomics, highlights good discriminative performance for predicting MVI in HCC, and may promote individualized therapeutic planning.