Abstract: Objective To develop a machine learning model based on computed tomography (CT) radiomics for predicting the risk of radiation proctitis in patients with locally advanced rectal cancer (LARC) after neoadjuvant chemoradiotherapy (nCRT). Methods Patients with LARC underwent nCRT at Zhejiang Cancer Hospital from December 2021 to December 2024 were retrospectively enrolled as the training set (n=326), while patients underwent nCRT from August 2022 to August 2024 were prospectively enrolled as the validation set (n=104). The region of interest in mesorectal region was manually delineated on pre⁃treatment CT images. Handcrafted features and deep learning features were extracted using Pyradiomics and a pre⁃trained ResNet18 model, respectively. Feature selection was performed using univariable Logistic regression, Spearman correlation analysis, least absolute shrinkage and selection operator regression, and recursive feature elimination. Logistic regression, support vector machine, light gradient boosting machine, and eXtreme gradient boosting were employed to develop independent radiomics models （Rad⁃model）, independent models based on deep learning features （DL⁃model）, and the combined model integrating both feature sets （DLRad⁃model）. The predictive performance of the models was evaluated using the area under the curve (AUC) of receiver operating characteristic. Results The incidence of grade 2 or higher radiation proctitis did not differ significantly between the training set and the validation set (36.81% vs 25.96%, P=0.056). Both the DL⁃model and the DLRad⁃model constructed based on eXtreme gradient boosting achieved AUC of 0.901 (95%CI: 0.874-0.927) and 0.918 (95%CI: 0.893-0.940), respectively, in the training set. In the validation set, these models achieved AUC of 0.747 (95%CI: 0.644-0.845) and 0.729 (95%CI: 0.620-0.829), respectively. Both of which were significantly higher than that of the Rad⁃model (all P<0.05). However,  the difference in AUC between the DL⁃model and the DLRad⁃model was not statistically significant (P>0.05). Conclusions Deep learning features demonstrate advantages in predicting the risk of radiation proctitis after nCRT in patients with LARC, and providing an effective tool for clinical risk stratification.

Key words: Locally advanced rectal cancer, Neoadjuvant chemoradiotherapy, Radiomics, Deep learning, Radiation proctitis