Prediction of overall survival in patients with colorectal cancer using baseline 18F-FDG PET/CT radiomics

Abstract To predict 3-year overall survival (OS) in patients with colorectal cancer (CRC) using baseline 18 F-FDG PET/CT radiomics. This retrospective study entrolled 166 patients with colorectal cancer (CRC) , who were randomly assigned into a training cohort ( n = 133) and an internal random hold-out validation cohort ( n = 33). The optimum subset was screened out from 2446 radiomic features via least absolute shrinkage and selection operator (LASSO) Cox regression. Significant clinical characteristics identified by Cox analysis were combined with radiomic score (Radscore) to construct a combined prediction model. The Akaike information criterion (AIC), Harrell’s concordance index (C-index) and the area under the receiver operating characteristic curve (AUC) were applied to evaluate model performance. The complex model had the lowest AIC, highest C-index with optimal efficacy across both cohorts. In the training cohort, its AUC was superior to the single radiomics (0.877 vs. 0.784, p = 0.027) and clinical model (0.877 vs. 0.774, p < 0.001), with no significant difference between the two alone (0.784 vs. 0.774, p = 0.878). In the internal validation cohort, its AUC differed significantly from the clinical model (0.896 vs. 0.822, p = 0.012). By contrast, no significant differences were found between the radiomics model and the complex model (0.707 vs. 0.896, p = 0.094), nor between the radiomics model and the clinical model (0.707 vs. 0.822, p = 0.400). Baseline PET/CT radiomics enables effective prediction of 3-year OS in patients with CRC. The complex model could further improve the predictive efficacy of the clinical model alone.