Objective:To investigate the diagnostic value of different conventional ultrasound-based radiomics models and their combination with clinical ultrasound features in the pathological subtyping of renal cell carcinoma.Methods:Retrospective data from 286 patients diagnosed with renal cell carcinoma by pathology at the Tenth Affiliated Hospital of Southern Medical University between May 1，2017 and June 7，2024 were collected. Among the 286 patients，203 were clear cell carcinoma，44 were papillary renal cell carcinoma，and 39 were chromophobe renal cell carcinoma. The patients were randomly divided into a training group（201 cases）and a validation group（85 cases）in a ratio of 7 to 3. Regions of interest（ROI）were delineated on conventional ultrasound images，and the radiomics features were extracted. Feature selection was performed using Student's t-test，Pearson correlation，and the least absolute shrinkage and selection operator（LASSO）. Six different machine learning methods included category gradient boosting（CatBoost），light gradient boosting machine（LightGBM），Logistic regression（LR），random forest（RF），support vector machine（SVM）and extreme gradient boosting（XGBoost）were used to establish radiomics models. Weight balancing was applied to correct for sample imbalance，and an imaging genomics model was constructed after balancing the samples. Independent predictors of renal cell carcinoma subtyping were selected from clinical ultrasound features using univariate and multivariate logistic regression analyses，and a clinical imaging model was constructed. The best-performing radiomics model was combined with the clinical independent predictors to construct a combined model. Receiver operating characteristic curves and the obuchowski index were plotted to evaluate model performance. Results:Among the radiomics models，the model constructed using Random Forest（RS RF）after balancing the samples exhibited the best predictive performance，with area under the curve（AUCs）of 0.918（micro-average ROC）and 0.903（macro-average ROC），and the obuchowski index was 0.885 in the validation group. The long and short axes of ultrasound image tumor masses were used as imaging independent predictors to construct a clinical imaging model. In the validation group，the AUCs of the clinical model were 0.886（micro-average ROC）and 0.606（macro-average ROC），and the obuchowski index was 0.569. The combined model achieved AUCs of 0.888（micro-average ROC）and 0.967（macro-average ROC），with an obuchowski index of 0.933，outperforming any single model. Conclusions:The combination of conventional ultrasound-based radiomics models with clinical ultrasound features demonstrates high diagnostic value in differentiating clear cell carcinoma，papillary renal cell carcinoma，and chromophobe renal cell carcinoma. It may serve as an auxiliary tool for providing timely and effective clinical guidance.

Objective:To investigate the diagnostic value of different conventional ultrasound-based radiomics models and their combination with clinical ultrasound features in the pathological subtyping of renal cell carcinoma.Methods:Retrospective data from 286 patients diagnosed with renal cell carcinoma by pathology at the Tenth Affiliated Hospital of Southern Medical University between May 1，2017 and June 7，2024 were collected. Among the 286 patients，203 were clear cell carcinoma，44 were papillary renal cell carcinoma，and 39 were chromophobe renal cell carcinoma. The patients were randomly divided into a training group（201 cases）and a validation group（85 cases）in a ratio of 7 to 3. Regions of interest（ROI）were delineated on conventional ultrasound images，and the radiomics features were extracted. Feature selection was performed using Student's t-test，Pearson correlation，and the least absolute shrinkage and selection operator（LASSO）. Six different machine learning methods included category gradient boosting（CatBoost），light gradient boosting machine（LightGBM），Logistic regression（LR），random forest（RF），support vector machine（SVM）and extreme gradient boosting（XGBoost）were used to establish radiomics models. Weight balancing was applied to correct for sample imbalance，and an imaging genomics model was constructed after balancing the samples. Independent predictors of renal cell carcinoma subtyping were selected from clinical ultrasound features using univariate and multivariate logistic regression analyses，and a clinical imaging model was constructed. The best-performing radiomics model was combined with the clinical independent predictors to construct a combined model. Receiver operating characteristic curves and the obuchowski index were plotted to evaluate model performance. Results:Among the radiomics models，the model constructed using Random Forest（RS RF）after balancing the samples exhibited the best predictive performance，with area under the curve（AUCs）of 0.918（micro-average ROC）and 0.903（macro-average ROC），and the obuchowski index was 0.885 in the validation group. The long and short axes of ultrasound image tumor masses were used as imaging independent predictors to construct a clinical imaging model. In the validation group，the AUCs of the clinical model were 0.886（micro-average ROC）and 0.606（macro-average ROC），and the obuchowski index was 0.569. The combined model achieved AUCs of 0.888（micro-average ROC）and 0.967（macro-average ROC），with an obuchowski index of 0.933，outperforming any single model. Conclusions:The combination of conventional ultrasound-based radiomics models with clinical ultrasound features demonstrates high diagnostic value in differentiating clear cell carcinoma，papillary renal cell carcinoma，and chromophobe renal cell carcinoma. It may serve as an auxiliary tool for providing timely and effective clinical guidance.