MRI-based radiomics model for predicting the nature of nodules in cirrhotic liver

To assess the value of magnetic resonance imaging (MRI)-based radiomics in predicting hepatocellular carcinoma (HCC), regenerative nodules (RNs) and dysplastic nodules (DNs). A total of 266 patients with pathologically confirmed RNs, DNs and HCCs were retrospectively analyzed. The dataset for model construction included 282 lesions (56 RNs,72 DNs and 154 HCCs) and the independent external validation dataset included 80 lesions (20 RNs, 20 DNs and 40 HCCs). All patients underwent MRI enhancement using T 1 -weighted imaging (T 1 WI), T 2 -weighted imaging (T 2 WI), Diffusion-weighted imaging (DWI) and T 1 WI-enhanced sequences. A total of 2264 radiomic features were extracted from the regions of interest. The best features were selected using five-fold cross-validation and a semantic feature-based model, a radiomics-based model and a combination model of semantic and radiomic features were established respectively with machine-learning techniques. These models were internally and externally validated. The decision curve analysis, calibration curves and area under the curve (AUC) were used to evaluate the model's performance. In the external validation, the performance of the combined model was the best, with an AUC of 0.896, which was higher than that of the radiomics-based model (AUC = 0.859) and semantic features-based model (AUC = 0.883). Decision curve analysis and the calibration curve demonstrated that the combined model was more useful in the external validation cohort. Machine learning based on a combined model (semantic and radiomic features) is significantly more advantageous in predicting RNs, DNs and HCCs than a semantic feature-based model or radiomics-based model alone.