Abstract
The purpose of this study was to construct a prognostic model of hepatocellular carcinoma (HCC) patients based on radiomics features and to explore the corresponding mechanisms. Patients' data and computed tomography images were collected from the Cancer Genome Atlas and the Cancer Imaging Archive. The radiomics features were extracted and screened via 3D Slicer and the Least Absolute Shrinkage and Selection Operator regression analysis for radiomics model construction, with rad score (RS) calculated. The net reclassification index, integrated discrimination improvement, receiver operating characteristic curve, and decision curve analysis were utilized to evaluate the prediction efficiency of the model. Restricted Cubic Splines and Kaplan-Meier were used to analyze the association between RS and overall survival (OS). The differentially expressed genes between the high- and low-RS groups were collected for enrichment analyses to screen the key pathways and core genes. Based on 4 key features associated with OS, a model was constructed: RS = 0.626* Gray Level Nonuniformity Normalized + 0.063*Dependence Nonuniformity Normalized - 0.253* Long Run Low Gray Level Emphasis - 35.93* Contrast. The radiomics model had superior performance in predicting OS, and high RS was associated with poor OS (P < .05). Moreover, RS may influence HCC prognosis through ADH1A regulation of the retinol metabolism pathway. The prognostic model based on radiomics features has good predictive performance in HCC patients. The RS model may promote HCC progression through ADH1A regulation of the retinol metabolism pathway.