Application Of CT Radiomics In Subtype Classification,Pathological Grading And Survival Prediction Of Renal Cell Carcinoma

(Background]The therapeutic plan and prognosis of renal cell carcinoma(RCC)are closely related to histological subtypes and nuclear grades,and the beneficiaries of postoperative adjuvant therapy are unknown.The existing prediction models of RCC lack imaging parameters,and the stable application of radiomics,which can reflect tumor heterogeneity,in RCC needs to be further explored.[Objective]To investigate the feasibility of stable and reliable prediction of subtypes of RCC and nuclear grading of clear cell RCC based on 3D CT radiomic features(RFs).Based on the feasibility,using the RFs to develop and verify the postoperative progressionfree survival(PFS)nomogram of clear cell RCC,in order to provide a reference for personalized risk assessment and precise treatment of RCC,and improve the prognosis of patients.[Materials and Methods]Preoperative CT images and clinical data of patients with RCC were retrospectively analyzed.Using ITK-SNAP software based on PyRadiomics computing platform,the tumor plain scan and enhanced CT 3D volume areas of interest were obtained by manual segmentation,and RFs were extracted.Cross-validation is used to ensure the stability and reliability of feature selection and prediction model,and the performance of the model is expressed by the average value.Part Ⅰ:based on the unbalanced data,Ensemble learning bagging algorithms feature selection algorithm framework and lasso regression,respectively using logistic regression stage construction of single model and the whole image group learning model,a total of five model,and compared with clinical feature model,stable and reliable exploration of clear cell RCC,chromophobe cell RCC and papillary cells RCC three categories and two classification prediction model.Part Ⅱ:The stability feature selection framework and lasso regression were used for feature selection,and logical regression was used to construct 15 single,double,third and full period RFs models respectively,and compared with clinical characteristics,the decision tree and support vector machines and random forests model contrast,to explore the optimal clear cell renal carcinoma WHO/ISUP nuclear grading forecast model.Part Ⅲ:The stability feature selection framework and lasso regression were used for feature selection.Survival analysis was conducted to explore the clinical-RFs clear cell RCC PFS nomogram,and its gain value was compared with clinical characteristic model.[Results]Part Ⅰ:The three-classification and two-classification prediction models of the whole period had the best performance,with a total accuracy of 0.80.The sensitivity of the test set to predict clear cell renal carcinoma,papillary cell renal carcinoma and chromophobe renal carcinoma was 0.85,0.60 and 0.66,respectively.The specificity was 0.83,0.91 and 0.91,respectively.AUC were 0.89,0.85 and 0.89,respectively.Part Ⅱ:Clinical-RFs combined model for predicting WHO/ISUP nuclear grading of clear cell renal carcinoma based on symptoms,maximum tumor diameter,plain CT value,and RFs was the best,with an AUC of 0.82 in the training set and 0.77 in the test set,respectively.Part Ⅲ:Clinical-radiomic nomograms based on age,clinical stage,KPS score,and RFs predict optimal clear cell renal carcinoma PFS performance.The C-index of training set and test set were 0.836 and 0.706,respectively.Decision curve analysis showed that the efficiency of the nomogram was always better than that of the clinical feature model,and the improvement rate of reclassification was 18.03%and the improvement rate of comprehensive discrimination was 19.08%.[Conclusion]The ensemble learning bagging algorithm can be used to predict RCC subtypes stably and reliably based on CT 3D radiomics model.The stability feature selection framework can be used to reliably predict WHO/ISUP nuclear grading of clear cell RCC based on plain CT clinical-RFs combined model.The use of clinical-RFs nomogram can better predict RCC PFS and optimize the efficacy of clinical characteristic model,which can provide a non-invasive reference for clinicians to individualize the risk assessment of renal cancer and formulate treatment plans.