Developing A Model To Predict Pathological Response To Neoadjuvant Chemoradiotherapy In Rectal Cancer Based On Radiomics Combined With Machine Learning

ObjectiveTo explore the influence of pathological response of neoadjuvant chemoradiotherapy(NCRT)on the prognosis of locally advanced rectal cancer(LARC),a novel model was established to predict pathological response based on the radiomics features of rectal magnetic resonance imaging(MRI)extraction before NCRT,and we explored the potential mechanism of radiomics features to predict pathological response,to provide a theoretical basis for subsequent studies.MethodsA total of 226 patients with locally advanced rectal cancer were collected from the Cancer Hospital of the Chinese Academy of Medical Sciences,Cox proportional risk regression was used to determine the independent factors of disease-free survival(DFS)and overall survival(OS).The Kaplan-Meier survival curve was used to estimate the 5year DFS and 5-year OS of rectal cancer patients.The log-rank test was used to compare survival differences.The radiomics features of rectal T2-weighted imaging(T2WI)extracted before NCRT were analyzed,and all patients were randomly divided into a training set and a validation set in a 7:3 ratio.The T-test(P<0.1)and the least absolute shrinkage and selection operator(LASSO)algorithm identify the radiomics features.The prediction ability of different machine learning(ML)algorithms was compared by receiver operator characteristic(ROC)curve and confusion matrix.Logistic regression(LR)was used to calculate the weight coefficients of each feature to estimate the radiomics score(Radscore),and to determine the independent factors of pathological response to neoadjuvant therapy.Develop a nomogram model to predict the pathological response of rectal cancer,calibration curve and decision curve analysis(DCA)to verify the accuracy and performance of the model.The immunohistochemical(IHC)experiment was conducted to investigate the expression of CD4,CD8,and HIF in rectal cancer specimens before NCRT.ResultsA total of 226 patients with locally advanced rectal cancer who received NCRT were included in this study,including 117 patients with good response and 109 patients with poor response.Cox proportional regression model showed that pathological response was an independent factor of DFS(poor response:HR,2.06;95%CI,1.04-4.07;P=0.037)and OS(poor response:HR,2.39;95%CI,1.05-5.43;P=0.038).The Kaplan-Meier survival analysis showed that the 5-year DFS of the good response group was significantly better than that of the poor response group(84.31%vs.70.00%,P=0.006).The 5-year OS of the good response group was significantly better than that of the poor response group(91.63%vs.74.82%,P=0.001).All patients were randomly divided into the training set and the validation set in a 7:3 ratio,of which 159 patients(70.3%)were in the training set and 67 patients(29.7%)were in the validation set.T-test and LASSO algorithm are used to determine 9 radiomics features.Six machine learning algorithms were used to construct radiomics models.In the training set,the area under curve(AUC)of the random forest(RF)model was 0.9999(95%CI:0.9996-1.0000),and the AUC of the extreme gradient boosting(XGBoost)model was 0.9998(95%CI:0.9993-1.0000),the AUC of support vector machine(SVM)was:0.9278(95%CI:0.8853-0.9703),the AUC of decision tree(DT)model was:0.9265(95%CI:0.8835-0.9695),the AUC of logistic regression(LR)model was:0.8771(95%CI:0.8235-0.9307),the AUC of naive bayesian(NB)model was:0.8411(95%CI:0.7800-0.9023);In the validation set,the AUC of RF model was:0.8884(95%CI:0.8001-0.9767),the AUC of XGBoost model was:0.8358(95%CI:0.7377-0.9338),the AUC of SVM model was:0.8276(95%CI:0.7221-0.9331),the AUC of DT model was:0.6652(95%CI:0.5301-0.8002),the AUC of LR model was:0.8267(95%CI:0.72640.9270),the AUC of NB model was:0.8058(95%CI:0.6969-0.9147);The RF algorithm performed best in the sensitivity,specificity,positive predictive value(PPV),negative predictive value(NPV)and F1 score of the predictive model.Logistic regression found CA199 status(abnormal CA199:HR,1.213;95%CI,1.036-1.420;P=0.018)and Radscore(HR,1.121;95%CI,1.093-1.150;P<0.001)are independent factors of pathological response,and radiomics model and nomogram model were established.The calibration curve showed good consistency and the clinical decision curve showed good clinical utility.The exploratory study found the mean optical density(MOD)of CD4(MOD:good response 2.558,poor response 0.873;P=0.001),CD8(MOD:good 0.828,poor 0.121;P=0.001)were significantly higher than that of patients with poor response,while HIF(MOD:good response 7.179,poor response 13.210;P<0.001)was significantly higher in patients with poor response,in rectal cancer tissue before neoadjuvant therapy.ConclusionsIn locally advanced rectal cancer,pathological response to neoadjuvant therapy is an independent influencing factor for DFS and OS.Patients with rectal cancer who obtained good response have better DFS and OS.Radiomics predict the pathological response of neoadjuvant therapy,and the random forest algorithm has the best performance and accuracy.The Radiomics model and Nomogram model developed by Rad-Score and CA199 have good prediction accuracy and clinical application value,which needs further study to verify.CD4,CD8,and HIF may be the potential mechanisms for imaging to predict pathological responses,which need to be verified by further basic experiments.