Objective:To explore the feasibility of radiomics-based quantitative analysis for molecular pathological subtyping in renal computed tomography angiography（CTA）and to establish a predictive model for renal mass subgroups.Methods:We retrospectively enrolled 535 patients with renal masses，including malignant lesions［223 clear cell renal cell carcinomas（ccRCC），84 papillary renal cell carcinomas（pRCC），113 chromophobe renal cell carcinomas（chrRCC）］and benign lesions［62 fat-poor angiomyolipomas（fpAML），53 oncocytomas］. There were 195 males and 340 females，with a median age of 52（range 49 to 80）years old. All patients underwent standard renal CTA prior to surgery. Radiomics features were extracted from CTA images. Data were categorized into six subgroups（malignant vs. benign，ccRCC vs. other renal masses，pRCC vs. other renal masses，chrRCC vs. other renal masses，fpAML vs. other renal masses，oncocytomas vs. other renal masses）. The dataset was randomised into training and validation cohorts by dividing the patients in a 2∶1 ratio. A machine learning-based predictive model（Radiomics-CTA）was developed using selected radiomic features in the training cohort. The model efficacy was assessed in the training cohort and validation cohort separately by plotting subject operating characteristic（ROC）curves，calculating area under the curve（AUC），and plotting clinical decision curves for model efficacy assessment.Results:For the malignant subgroup，Radiomics-CTA achieved area under the receiver operating characteristic curve（AUC）values of 0.823（95% CI 0.751?0.894）and 0.833（95% CI 0.783?0.883）in the training and validation cohorts，respectively. For ccRCC identification，the model showed AUCs of 0.928（95% CI 0.89?0.955）and 0.925（95% CI 0.881?0.968）in the two cohorts. For the other subtypes identification，such as pRCC，chrRCC，fpAML，and oncocytomas，the model showed AUCs of 0.862（95% CI 0.826?0.898），0.882（95% CI 0.849?0.915），0.921（95% CI 0.898? 0.943），and 0.865（95% CI 0.787?0.944）in the training cohort，and the AUC of 0.823（95% CI 0.776?0.870），0.842（95% CI 0.754?0.929），0.930（95% CI 0.892?0.968）and 0.876（95% CI 0.847? 0.906）in the validation cohort . Radiomics-CTA outperformed senior radiologists in diagnosing ccRCC［87.1%（466/535）vs. 83.2%（445/535）， P=0.03）］and chrRCC［82.1%（439/535）vs. 80.0（428/535）， P<0.01］. Conclusions:The Radiomics-CTA model can extract deep pathological information from CTA images through radiomics methods，and has the ability to distinguish pathological subtypes of renal tumors. It can also provide assistance for accurate diagnosis by radiologists to a certain extent.

Objective:To explore the feasibility of radiomics-based quantitative analysis for molecular pathological subtyping in renal computed tomography angiography（CTA）and to establish a predictive model for renal mass subgroups.Methods:We retrospectively enrolled 535 patients with renal masses，including malignant lesions［223 clear cell renal cell carcinomas（ccRCC），84 papillary renal cell carcinomas（pRCC），113 chromophobe renal cell carcinomas（chrRCC）］and benign lesions［62 fat-poor angiomyolipomas（fpAML），53 oncocytomas］. There were 195 males and 340 females，with a median age of 52（range 49 to 80）years old. All patients underwent standard renal CTA prior to surgery. Radiomics features were extracted from CTA images. Data were categorized into six subgroups（malignant vs. benign，ccRCC vs. other renal masses，pRCC vs. other renal masses，chrRCC vs. other renal masses，fpAML vs. other renal masses，oncocytomas vs. other renal masses）. The dataset was randomised into training and validation cohorts by dividing the patients in a 2∶1 ratio. A machine learning-based predictive model（Radiomics-CTA）was developed using selected radiomic features in the training cohort. The model efficacy was assessed in the training cohort and validation cohort separately by plotting subject operating characteristic（ROC）curves，calculating area under the curve（AUC），and plotting clinical decision curves for model efficacy assessment.Results:For the malignant subgroup，Radiomics-CTA achieved area under the receiver operating characteristic curve（AUC）values of 0.823（95% CI 0.751?0.894）and 0.833（95% CI 0.783?0.883）in the training and validation cohorts，respectively. For ccRCC identification，the model showed AUCs of 0.928（95% CI 0.89?0.955）and 0.925（95% CI 0.881?0.968）in the two cohorts. For the other subtypes identification，such as pRCC，chrRCC，fpAML，and oncocytomas，the model showed AUCs of 0.862（95% CI 0.826?0.898），0.882（95% CI 0.849?0.915），0.921（95% CI 0.898? 0.943），and 0.865（95% CI 0.787?0.944）in the training cohort，and the AUC of 0.823（95% CI 0.776?0.870），0.842（95% CI 0.754?0.929），0.930（95% CI 0.892?0.968）and 0.876（95% CI 0.847? 0.906）in the validation cohort . Radiomics-CTA outperformed senior radiologists in diagnosing ccRCC［87.1%（466/535）vs. 83.2%（445/535）， P=0.03）］and chrRCC［82.1%（439/535）vs. 80.0（428/535）， P<0.01］. Conclusions:The Radiomics-CTA model can extract deep pathological information from CTA images through radiomics methods，and has the ability to distinguish pathological subtypes of renal tumors. It can also provide assistance for accurate diagnosis by radiologists to a certain extent.