Objective: Rectal adenoma is a precancerous lesion of rectal cancer(RC),and it is difficult to differentiate from T1-2 stage RC,which is confined to the intestinal wall invasion,on medical imaging,leading to misdiagnosis and mistreatment.In this study,radiomics models based on pre-operative contrast-enhanced computed tomography(CECT)were developed and validated to identify T1-2 stage RC and adenoma,and the performance of the models with that of radiologists were compared.The aim of this study is to provide new ideas and methods for the precise diagnosis and treatment of T1-2 stage RC and adenoma.Methods: Clinical,pathologic and imaging data were retrospectively collected from patients who underwent surgery and were confirmed by postoperative pathology as rectal adenomas or p T1-2 adenocarcinomas from June 2014 to December 2021 at Sir Run Run Shaw Hospital.All patients underwent CECT examination before surgery.The entire tumor area was manually outlined layer by layer on the CECT venous phase images,and imaging features were extracted.The optimal predictive radiomics features were selected through the Least Absolute Shrinkage and Selection Operator(LASSO)algorithm,while the clinical predictive factors were obtained through multifactorial logistic analysis.Nine models were constructed based on the radiomics features,clinical factors,and their combination using three machine learning algorithms,namely Random Forest(RF),Support Vector Machine(SVM),and Logistic Regression(LR).The models were trained and validated using five-fold cross-validation.The performance of the models was evaluated using Receiver Operating Characteristic(ROC)curves,calibration curves,and Decision Curve Analysis(DCA).To compare with clinical diagnostic accuracy,two experienced radiologists classified each patient separately.Finally,to further investigate the performance of the models,a specific analysis of correctly and incorrectly classified cases was performed.Results: A total of 311 patients(males,58.5%;mean age,63.18 ± 9.85 years)were included,including 106 adenomas(males,57.5%;mean age,63.11 ± 10.20 years)and 205 p T1-2 adenocarcinomas(males,59.0%;mean age,63.21 ± 9.69 years).Eight imaging features and two clinical predictors were selected as the optimal predictive features to build the model.Among the three classifiers,the RF classifier had the best predictive performance.In the model based on imaging features alone,the RF algorithm achieved an accuracy of 86% and an Area under curve(AUC)of 0.90,while the model combining clinical and imaging features showed higher predictive performance with the same accuracy,with an AUC of 0.93.The calibration curve shows that the predicted values of the best model are consistent with the actual observed values,and DCA indicates that the model has good clinical application value.In contrast,the two radiologists achieved accuracies of 62% and 68%,respectively,with AUCs of 0.60 and 0.70.The diagnostic performance of the best model was influenced by the pathological classification and staging: there was higher accuracy(about 90.3%)in classifying low-grade intraepithelial neoplasia adenomas;well-differentiated adenocarcinomas had a higher probability(28.6%)of being classified as adenomas compared to moderately differentiated adenocarcinomas(17.1%);the probability of misclassifying T1-stage RC was twice as high as that of T2-stage(about 29%/14.7%).Conclusion: 1.The CECT-based radiomics model can be used to differentiate between T1-2 stage RC and adenoma before surgery and is superior to experienced radiologists.2.The combined model of radiomics features and clinical characteristics performs better than models that use only radiomics features or clinical factors alone.3.Among the three machine learning algorithms,RF performs better than SVM and LR in the classification of T1-2 stage RC and adenoma,while SVM and LR show similar performance. |