Research On CT-based Radiomcis For Predicting WHO/ISUP Grading Of Clear Cell Renal Cell Carcinoma

[Background]Clear cell renal cell carcinoma(ccRCC)is a subtype of renal cell carcinoma(RCC)with poor prognosis.The WHO/ISUP nuclear grading of ccRCC is significantly correlated with the 5-year survival rate of patients,which is an independent factor in predicting the prognosis of patients.Targeted treatment of patients with different nuclear grades may effectively improve the survival time and quality of life of patients.At present,histopathology is mainly used to detect the nuclear grade of ccRCC tumor in clinical practice.However,patients with advanced stage have poor tolerance,and there are still inconsistencies between biopsy and gross pathology.At the same time,there are still some limitations depending on the clinical experience and operation level of clinicians.Radiomics can effectively improve the above problems by reflecting the heterogeneity of tumors through its high-throughput radiomics features.However,most of the feature selection methods in radiomics usually have some problems,such as poor robustness of prediction model and insufficient interpretability of selection features.Therefore,a differential network analysis feature selection method is introduced in this paper to select the radiomics features of CT renal carcinoma.and the feasibility of constructing a nuclear grade prediction model based on this method is studied in this paper.[Objective]1.The WHO/ISUP nuclear grading prediction model of CT radiomics was established based on differential network analysis.2.The correlation between the radiomics features of the WHO/ISUP nuclear grading prediction model and prognosis of patient was investigated by survival analysis.[Materials and methods]A retrospective analysis was conducted on 175 patients(105 in the training set and 70 in the test set)admitted to Nanfang Hospital of Southern Medical University from 2015 to 2021.The lesions were delineate by layer on the non-contrast phase(NCP)and three enhanced phases of CT images to form the three-dimensional volume of interest(VOI).ITK-SNAP and PyRadiomics software platforms were used for segmentation and feature extraction of tumor volume images,and 107 RFs were extracted for each phase.Chapter Ⅱ:In the training set,feature selection is made for each phase through differential network analysis to obtain their respective radiomics features.Then,Logistic regression was used to construct the WHO/ISUP nuclear grading prediction model for different phases.The phase model with the best performance was selected and compared with other machine learning and clinical models through five ten-fold layered cross-validations.Finally,the ROC and AUC were used for model evaluation,and the DeLong test was used to compare results.Chapter Ⅲ:Testing the correlation between the radiomics features in the optimal WHO/ISUP nuclear grading prediction model and prognosis of patient by correlation survival analysis.Kaplan-Meier survival analysis,iunivariate and multivariate Cox proportional hazard regression models were used to study the radiomics features associated with progression-free survival.Finally,risk scores were used to visualize the retained radiomics features.[Results]Chapter Ⅱ:1.Comparison of single-phase and whole-phase radiomics models:The pred ictive efficiency of NCP model was the best and the AUC of the training and validation sets were 0.78 and 0.76,respectively.2.Comparison between the NCP model and different machine learning models:In the training set,NCP model and RF model were better than other models,but there was no statistical difference between them(p>0.05).In the test set,the performance of most machine learning models also decreased significantly while the NCP model remained stable.3.Comparison between the NCP model and the clinical model:In both the validation set and the test set,the NCP model was the best(AUC was 0.76 and 0.75,respectively),significantly superior to other models(p<0.05).Chapter Ⅲ:1.K-M survival analysis:Except for radiomics features X2(uniformity)(p=0.23),there were significant differences of all other features.2.Cox regression analysis:All other features were significant(p<0.05)and were risk factors for progression-free survival except for radiomics features X2,which were included in multivariate Cox regression analysis;Multivariate results showed that X3(Gray Level Non Uniformity,GLN)and X6(Gray Level Non Uniformity Normalized,GLNN)were still significant(p<0.05)and were independent risk factors for progression-free survival.3.Risk score analysis:Radiomics features in the NCP model can effectively predict patients’ progression-free survival.[Conclusion]1.The NCP model constructed by the radiomics features based on differential network analysis can effectively predict the WHO/ISUP nuclear grading of ccRCC.2.The radiomics features in the WHO/ISUP nuclear grading prediction model are associated with the prognosis of patients of ccRCC.They can reasonably predict the progression-free survival of patients with solid explanatory ability.