A pre-treatment MRI-based radiomics approach combinedwith clinical characteristics for predicting early recurrence in hepatocellular carcinoma after radiofrequency ablation treatment

Abstract Purpose To develop a prediction model combining MRI-based radiomic features with clinical indicators in predicting early recurrence of hepatocellular carcinoma (HCC) after radiofrequency ablation (RFA) treatment. Materials and Methods In total, 169 HCC patients who underwent MRI before RFA treatment from January 2015 to December 2021 were retrospectively enrolled at our institution. Patients were randomly divided into training cohorts (n=135) and test cohorts (n= 34). In the training cohorts, feature selection was performed by variance threshold, select K-best, and least absolute shrinkage and selection operator (LASSO).Then three machine-learning classifiers [support vector machine (SVM), logistic regression (LR), and random forest (RF)] were applied to build predictive models using radiomics or integrated radiomics-clinical characteristics. The models were tested in training cohorts and test cohorts according to the early recurrence of HCC by the receiver operating characteristic curve (ROC) analysis. Results Alpha fetoprotein levels, platelet count, and the location of HCC were effective clinical indicators for predicting early recurrence of HCC after treatment with RFA. Progressive reduction of MRI radiomics features by dimension screening identified 16 effective features. The predictive models based on radiomics-only showed good performance across all classifiers constructed by SVM, LR, and RF (SVM, train: AUC 0.957 [0.920-0.993], test: AUC 0.826 [0.677-0.925]; LR, train: AUC 0.956 [0.924-0.987], test: AUC 0.830[0.674-0.986]; RF, train: AUC 0.945 [0.909-0.982], test: AUC 0.826 [0.682-0.969]). The radiomics-clinical integrated predictive models (clinical, routine MRI image features, and radiomics parameters) showed better performance by SVM, LR, and RF classifiers (SVM, train: AUC 0.956 [0.921-0.992], test: AUC 0.830 [0.686-0.973]; LR, train: AUC 0.957 [0.926-0.988], test: AUC 0.830[0.673-0.986]; RF, train: AUC 0.975 [0.952-0.997], test: AUC 0.909 [0.807-1.000]), in which the RF classifier performed best (0.909 vs 0.830 vs 0.830, P＜0.05) . Conclusion The MRI radiomics-clinical integrated predictive models effectively predicted early recurrence in HCC patients treated with RFA, and the RF classifier may be the most effective model.