Predicting pathological complete response in rectal cancer after chemoradiotherapy with a random forest using (18)F-fluorodeoxyglucose positron emission tomography and computed tomography radiomics

BACKGROUND: Neoadjuvant chemoradiotherapy (NCRT) followed by surgery is the standard treatment for patients with locally advanced rectal cancer. This study developed a random forest (RF) model to predict pathological complete response (pCR) based on radiomics derived from baseline (18)F-fluorodeoxyglucose ([(18)F]FDG)-positron emission tomography (PET)/computed tomography (CT). METHODS: This study included 169 patients with newly diagnosed rectal cancer. All patients received (18)F[FDG]-PET/CT, NCRT, and surgery. In total, 68 radiomic features were extracted from the metabolic tumor volume. The numbers of splits in a decision tree and trees in an RF were determined based on their effects on predictive performance. Receiver operating characteristic curve analysis was performed to evaluate predictive performance and ascertain the optimal threshold for maximizing prediction accuracy. RESULTS: After NCRT, 22 patients (13%) achieved pCR, and 42 features that could differentiate tumors with pCR were used to construct the RF model. Six decision trees and seven splits were suitable. Accordingly, the sensitivity, specificity, positive predictive value, negative predictive value, and accuracy were 81.8%, 97.3%, 81.8%, 97.3%, and 95.3%, respectively. CONCLUSIONS: By using an RF, we determined that radiomics derived from baseline (18)F[FDG]-PET/CT could accurately predict pCR in patients with rectal cancer. Highly accurate and predictive values can be achieved but should be externally validated.