MRI-based radiomics model for preoperative prediction of 5-year survival in patients with hepatocellular carcinoma

Abstract

Background: Recurrence is the major cause of mortality in patients with resected HCC. However, without a standard approach to evaluate prognosis, it is difficult to select candidates for additional therapy.

Methods: A total of 201 patients with HCC who were followed up for at least 5 years after curative hepatectomy were enrolled in this retrospective, multicentre study. A total of 3144 radiomics features were extracted from preoperative MRI. The random forest method was used for radiomics signature building, and five-fold cross-validation was applied. A radiomics model incorporating the radiomics signature and clinical risk factors was developed.

Results: Patients were divided into survivor (n = 97) and non-survivor (n = 104) groups based on the 5-year survival after surgery. The 30 most survival-related radiomics features were selected for the radiomics signature. Preoperative AFP and AST were integrated into the model as independent clinical risk factors. The model demonstrated good calibration and satisfactory discrimination, with a mean AUC of 0.9804 and 0.7578 in the training and validation sets, respectively.

Conclusions: This radiomics model is a valid method to predict 5-year survival in patients with HCC and may be used to identify patients for clinical trials of perioperative therapies and for additional surveillance.

Fig. 1. Study recruitment process.

HCC hepatocellular carcinoma, MRI magnetic resonance imaging.

Fig. 2. Workflow of necessary steps in this study.

The region of interest (ROI) in each transverse section was semi-automatically segmented on T1-weighted, T2-weighted, diffusion-weighted and dynamic contrast-enhanced magnetic resonance images. After three-dimensional reconstruction of the ROI, 3144 features, including 786 for each sequence, were extracted, and the top 30 were selected via Gini coefficient. Based on the selected features and clinical risk factors, a radiomics model was developed using the random forest method and five-fold cross-validation. The performance of the radiomics model was evaluated according to receiver-operating characteristic, calibration and decision curves. HOG histogram of oriented gradient, ROC receiver-operating characteristic.

Fig. 3. Receiver-operating characteristic (ROC) curves for the radiomics signature and radiomics model.

a ROC curve for the radiomics signature in the training set, showing a mean area under the curve (AUC) of 0.9733 (95% confidence interval [CI], 0.9671–0.9795). b ROC curve for the radiomics signature in the validation set, showing a mean AUC of 0.7025 (95% CI, 0.6695–0.7355). c ROC curve for the radiomics model with the addition of preoperative alpha-fetoprotein (AFP) and aspartate aminotransferase (AST) in the training set, showing a mean AUC of 0.9804 (95% CI, 0.9714–0.9894). d ROC curve for the radiomics model with the addition of preoperative AFP and AST in the validation set, showing a mean AUC of 0.7578 (95% CI, 0.7056–0.8100).

Fig. 4. Calibration curves for the radiomics model.

Calibration curves for the radiomics model in the training (a) and validation (b) sets. The diagonal blue line represents the perfect performance of an ideal model. The red, orange, green, purple and brown lines represent the performance of the radiomics model in five different training or validation sets, of which a closer fit to the diagonal blue line indicates a better prediction performance.