| BackgroundFor patients with locally advanced esophageal squamous cell carcinoma(ESCC),neoadjuvant immunochemotherapy has been shown to improve long-term outcomes,but the treatment response varies among patients.Accurate pretreatment prediction of response remains an urgent need.To develop radiomics models for predicting pathological complete response(PCR)to neoadjuvant immunochemotherapy in patients with esophageal squamous cell carcinoma by using different kinds of machine learning algorithms and explore the optimal model.MethodsIn this retrospective study,43 follow-up PET-CT images obtained in 43 patients with ESCC(mean age 60.91±8.1;31 men,12 women)were evaluated.According to the pathological response after surgery,the patients are divided into two groups:PCR group and non-PCR group.We extracted radiomic features of tumor on PET-CT images,and used Pearson correlation coefficient for feature selection,then 7 classification models including Extra Trees,Gradient Boosting,KNN,random forest(RF)and support vector machine(SVM),XG Boost were used to build predictive models and five-fold cross-validation assessing the accuracy.The performance of the models was evaluated based on Receiver operating characteristic curve(ROC)curves analyses and decision curve analysis(DCA).Results43 PET-CT images were registered in our study.A total of 107 radiomics features were extracted,of which 23 robust features are included in building the model.The consistency and effectiveness of model performance were compared and verified by different machine learning methods in different five cross-validation strategies,thus demonstrating the stability and robustness.Gradient Boosting classifier algorithm had the best prediction performance with an accuracy of 0.917 in the cohorts.ConclusionsThe results demonstrated that the PET-CT radiomics-based machine learning model has potential in predicting the PCR of ESCC. |
PDF Full Text Request