Panpan Jiao, Bin Wang, Xinmiao Ni, Yi Lu, Rui Yang, Yunxun Liu, Jingsong Wang, Haonan Mei, Xiuheng Liu, Xiaodong Weng, Qingyuan Zheng, Zhiyuan Chen
J Cancer . 2025 Jan 6;16(4):1118-1126. doi: 10.7150/jca.105173. eCollection 2025.
Purpose: Exploring the value of predicting the WHO/ISUP grade of clear cell renal cell carcinoma (ccRCC) using computed tomography urography (CTU) images, providing valuable recommendations for the treatment of ccRCC.
Method: CTU images from the Renmin Hospital of Wuhan University (RHWU) cohort, including 328 patients with ccRCC, were retrospectively collected. The corticomedullary (CMP) phase features of ccRCC were extracted from the CTU images using the Pyradiomics package, and key features were selected through the Least Absolute Shrinkage and Selection Operator (LASSO) regression. The 328 patients were split into training and testing sets in a 7:3 ratio. 175 patients from the The Cancer Genome Atlas (TCGA) cohort were used for the external validation set. Various models, including Logistic Regression (LR), Multilayer Perceptron (MLP), Support Vector Machine (SVM), and eXtreme Gradient Boosting (XGBoost), were employed to predict the ISUP grade. SHAP analysis was then used to visualize the performance of the best model.
Results: A total of 1,218 features were extracted using the Pyradiomics package, with 20 features selected for model training through LASSO analysis. In the training set, the AUC for the LR model was 0.88 (95% confidence interval [CI] 0.84-0.91), for MLP it was 0.89 (95% CI 0.86-0.93), for SVM it was 0.86 (95% CI 0.83-0.90), and for XGBoost it was 0.96 (95% CI 0.92-0.99). In the testing set, the AUC for LR was 0.79 (95% CI 0.73-0.85), for MLP it was 0.78 (95% CI 0.72-0.83), for SVM it was 0.78 (95% CI 0.73-0.82), and for XGBoost it was 0.80 (95% CI 0.75-0.85). In the validation set, the AUC for LR was 0.74 (95% CI 0.68-0.79), for MLP it was 0.68 (95% CI 0.63-0.73), for SVM it was 0.67 (95% CI 0.64-0.71), and for XGBoost it was 0.78 (95% CI 0.74-0.83). XGBoost demonstrated superior performance, with a sensitivity of 0.99 (95% CI 0.96-1.00) in the training set, 0.92 (95% CI 0.88-0.97) in the testing set and 0.91 (95% CI 0.86,0.95) in validation set. SHAP analysis revealed that the wavelet-LHL_glcm_Idn and wavelet-LHL_glrlm_LongRunEmphasis features played pivotal roles in the classification task.
Conclusion: In this study, we employ an artificial intelligence model to conduct non-invasive ISUP grade prediction on preoperative CTU images of ccRCC, thereby aiding clinical decision-making. Additionally, we uncover that the radiomics features extracted from the CMP phase of CTU images hold promise as potential biomarkers for grading ccRCC.