Radiomics-based decision support tool with ground-glass opacity status of 5-year survival prediction for early-stage non-small cell lung cancer.

Background: To evaluate the predictive performance differences for postoperative 5-year survival risk for early-stage non-small cell lung cancer (NSCLC) patients with two image analysis methods based on ground-glass opacity (GGO) status and radiomics analysis. Moreover, we improve the accuracy of stratifying survival risk by combining radiomics with GGO status. Materials and methods: Computed tomography (CT) images for 113 NSCLC patients were analyzed. The patients were divided into four groups according to %GGO step by 25%. The GGO model was built with the optimal cutoff %GGO value to categorize patients into high-risk or low-risk groups. The radiomics features were selected by the least absolute shrinkage and selection operator (LASSO)-Cox regression and these were incorporated into the Rad-score model. The combined model was created by integrating the GGO and Rad-score models. The survival rates between these groups were compared using Kaplan-Meier analysis, supplemented by log-rank tests. Results: From LASSO-Cox regression, 5 features were selected. Multivariate Cox regression analysis in the Combined model identified GGO and Rad-score as independent predictive factors. The combined model (C-index: 0.664) performed best compared to the GGO model (C-index: 0.521) and the rad-score model (C-index: 0.642). The 5-year survival Kaplan-Meier curves for the rad-score and combined models were also able to stratify the patient population into low-risk and high-risk groups (p-values < 0.05). Conclusion: A combined model, integrating the GGO status and Rad-score may help predict the prognosis of patients with early NSCLC more accurately, with a higher probability of outcome than a GGO model.