MR Radiomics Combined With Radiologic Features to Predict Recurrence Location in Nonviable Hepatocellular Carcinoma...

To develop a predictive model that integrates radiomics features from contrast-enhanced MRI with conventional radiologic features to identify early recurrence locations in nonviable hepatocellular carcinoma (HCC) after transarterial chemoembolization (TACE).

This multicenter retrospective study included HCC patients treated with TACE who were assessed as Liver Imaging Reporting and Data System Treatment Response Algorithm nonviable. All patients were followed for at least 1 year. A 1-cm peritumoral ring was divided into eight sectors for radiomics feature extraction to build a radiomics model. A fusion model was developed by combining radiomics features with two radiologic features (nonsmooth margin and peritumoral hyperintensity on T2-weighted imaging/diffusion-weighted imaging). Model performance was evaluated using receiver operating characteristic (ROC) curves. The DeLong test assessed differences in predictive performance.

The study finally included 424 sectors from 53 patients and 192 sectors from 24 patients in the training and test cohort, respectively. The radiomics model achieved an area under the ROC curve (AUC) of 0.771 and 0.602 in the training and test cohorts, respectively. The radiologic model achieved AUCs of 0.787 and 0.736 in the training and test cohorts, respectively. The fusion model combining six radiomics features and two radiologic features achieved AUC of 0.843 and 0.774 in the training and test cohorts, respectively. The DeLong test showed that the fusion model outperformed the radiomics and radiologic models in the training cohort and was superior to the radiomics model in the test cohort (p < 0.05).

The fusion model combining radiomics and radiologic features shows good performance in predicting recurrence location and may support personalized follow-up and retreatment planning.