Machine learning-based radiomics for preoperative prediction of microvascular invasion in hepatocellular carcinoma

To evaluate the diagnostic performance of different radiomics models for preoperatively predicting microvascular invasion (MVI) of hepatocellular carcinoma (HCC).

A total of 124 patients who had histologically confirmed HCC (training dataset: n=86; validation dataset: n=38) were included. Clinical factors (CFs) were extracted from medical data. Radiomics features were derived from the unenhanced phase, artery phase (AP), portal venous phase and delay phase CT images. The least absolute shrinkage and selection operator (LASSO) method was chosen to select the radiomics feature. Twelve models were established using three modeling methods (logistic regression [LR], support vector machine [SVM], and Bayes) with the radiomics signature. Receiver operating characteristic (ROC) analysis was used to evaluate the diagnostic performance of the models. A radiomics signature that performs well was integrated with the clinical factors into a combined model.

AP radiomics signatures achieved the best efficiency than other radiomics models. The LR model with the AP radiomics signature as the input factor showed better performance than the SVM and Bayes models, with an area under the curve (AUC) of 0.848 in the validation datasets. When integrating the AP radiomics signatures and clinical factors, the combined model performed better and reached an AUC of 0.875 in the validation datasets.

The radiomics model demonstrated excellent performance for preoperatively predicting MVI in patients with HCC, especially the combined model. Different modeling methods could influence the effects of the diagnostic performance.