Application Of Different Imaging Omics Models Based On Enhanced CT In Differentiating Renal Clear Cell Carcinoma Of Poor Blood Supply From Non-clear Cell Carcinoma

Objective:Contrast-enhanced CT was used to compare the diagnostic efficacy of imaging omics models based on different machine learning algorithms in renal clear cell carcinoma of poor blood supply and non-clear cell carcinoma.Materials and methods:This study was a retrospective study,and a total of 70 patients with renal malignancy were collected,including 37 cases of renal clear cell carcinoma and 33 cases of non-transparent renal carcinoma.The imaging data and clinicopathological information of all patients were complete.（1）Clinical data of each patient（age,sex,smoking,or not）were collected for statistical analysis.（2）this study enhanced CT images of 70 cases were imported to Dr.Wise’s multimodal research platform（https://keyan.deepwise.com）,which was made up of two practitioners of lesions of the CT image enhance image（including arterial and venous phase and delayed phase）manual step by step a sketch to extract the interesting area（ROI）,Image omics method was used to extract image omics features（first-order features,morphological features,texture features and higher-order features）,feature correlation analysis was used to reduce the dimensionality of these features,and the best phase was selected based on AUC values and accuracy of five different machine learning classifier models.Then,four feature selection methods and five machine learning classifiers were combined in pairs based on this time image,and 20 classification models were constructed through the tenfold cross test.The prediction performance of all models was evaluated by using the area under the curve（ROC）and accuracy,and the AUC heat maps of each classification model on the verification set were made.Results:There was no significant relationship between age,sex,smoking status,lesion location and subtype classification of renal carcinoma.In the three-phase enhanced models with different classifiers,the arterial phase had the best predictive performance,and the LR validation set AUC was 0.893 in the arterial phase.The accuracy was 0.828;The AUC of RF validation set was 0.897.The accuracy was0.828;The AUC of SVM validation set was 0.875.The accuracy was 0.857,and the KNN verification set AUC was 0.872.The accuracy is 0.757;The AUC of DT validation set is 0.850.The accuracy is 0.785.On this basis,four feature selection methods and five machine learning classifiers are combined in pairs based on the arterial phase.Among the 20 classification models,recursive feature elimination is the best.The AUC on the validation set of the "RFE-RF" model combined with the Random Forest（RF）classifier was 0.919.The accuracy was 0.842;The sensitivity was 0.888.The specificity was 0.794.The positive predictive value was 0.820.The negative predictive value was 0.871.Conclusion:1.Renal clear carcinoma of poor blood supply and non-clear renal cell carcinoma can be distinguished better in the arterial stage than in the venous stage and delayed stage.2.The RFE-RF imaging model based on the arterial phase of enhanced CT has a good performance in differentiating renal clear cell carcinoma of poor blood supply from non-clear cell carcinoma.