| Objective: To establish a prediction model of pathological grading of bladder cancer based on dynamic computed tomography（CT）radiomics,so as to achieve accurate preoperative prediction of pathological grading of bladder cancer.Methods: A total of 81 patients with bladder cancer in the hospital were selected for preoperative dynamic CT,including two types of imaging data: arterial and venous items.The postoperative cases of the patients were all diagnosed as non-invasive urothelial carcinoma.The outline of the bladder tumor is drawn,and the first-order feature variables,texture analysis variables,shape variables,wavelet transformation variables,990 feature variables are extracted,and a total of 1980 feature variables for each patient.The maxrelevance and min-redundancy（m RMR）and the least absolute shrinkage and selection operator（LASSO）are used for feature selection.Finally,a variety of machine learning algorithms are combined with meaningful feature variables to establish a predictive model for comparison of non-invasiveness The sensitivity,specificity and accuracy of the prediction of pathological grade of bladder cancer.Results: The radiomics features of the arterial and venous phases were used to obtain VOI,and feature variables were extracted;m RMR combined with LASSO was used to select 19 feature variables,and then four machine learning algorithms of KNN,NNET,RF,and SVM were used to establish and verify the model.The results of the four models are relatively consistent.Among them,the RF model is relatively stable,with an accuracy of 0.973（95%CI: 0.945-1.000）in the test set,and an accuracy of 0.601（95%CI: 0.354-0.848）in the verification set,and the comprehensive evaluation is also slightly better than other models.Conclusion: Preoperative dynamic CT radiomics strategy combined with machine learning algorithm can accurately predict the pathological grade of non-muscular invasive bladder cancer,and it is of scientific significance to promote personalized treatment of bladder cancer. |
