Predicting prognosis of resected hepatocellular carcinoma by radiomics analysis with random survival forest

To investigate the impact of random survival forest (RSF) classifier trained by radiomics features over the prediction of the overall survival of patients with resectable hepatocellular carcinoma (HCC).

The dynamic computed tomography data of 127 patients (97 men, 30 women; mean age, 68 years) newly diagnosed with resectable HCC were retrospectively analyzed. After manually setting the region of interest to include the tumor within the slice at its maximum diameter, texture analyses were performed with or without a Laplacian of Gaussian filter. Using the extracted 96 histogram based texture features, RSFs were trained using 5-fold cross-validation to predict the individual risk for each patient on disease free survival (DFS) and overall survival (OS). The associations between individual risk and DFS or OS were evaluated using Kaplan-Meier analysis. The effects of the predicted individual risk and clinical variables upon OS were analyzed using a multivariate Cox proportional hazards model.

Among the 96 histogram based texture features, RSF extracted 8 of high importance for DFS and 15 for OS. The RSF trained by these features distinguished two patient groups with high and low predicted individual risk (P=1.1×10^-4 for DFS, 4.8×10^-7 for OS). Based on the multivariate Cox proportional hazards model, high predicted individual risk (hazard ratio=1.06 per 1% increase, P=8.4×10^-8) and vascular invasion (hazard ratio=1.74, P=0.039) were the only unfavorable prognostic factors.

The combination of radiomics analysis and RSF might be useful in predicting the prognosis of patients with resectable HCC.