CT-based deep learning for preoperative prediction of pathological grading of renal clear cell carcinoma.

Objective: To construct a noninvasive preoperative prediction model for WHO/ISUP grading of renal clear cell carcinoma (ccRCC) using deep learning combined with four-phase CT images, and to evaluate its efficacy. Methods: A retrospective study was conducted on 158 ccRCC patients (124 low-grade, 34 high-grade) from the Affiliated Hospital of Hebei University (January 2022-June 2024). Patients were randomly divided into training, validation, and test sets at an 8:1:1 ratio. Four-phase CT images were preprocessed (rectangular box annotation of tumor region of interest [ROI], image resizing to 224×224 pixels). The ResNet34 model was first built to predict ccRCC grading, with performance evaluated by accuracy (ACC) and area under the receiver operating characteristic curve (AUC). The model was then optimized by integrating the SENet attention mechanism (forming the SE-ResNet34 model), and performance before and after optimization was compared. Results: ResNet34 models based on corticomedullary, parenchymal, and excretory phase images achieved ACC >0.8, with the parenchymal phase model showing the best performance (ACC = 0.867, low-grade AUC = 0.857, high-grade AUC = 0.853). After adding the SENet attention mechanism, the SE-ResNet34 model exhibited improved performance: ACC increased to 0.878, low-grade AUC to 0.929, and high-grade AUC to 0.927. Conclusion: The SE-ResNet34 model based on parenchymal phase CT images has excellent ability to differentiate WHO/ISUP grades of ccRCC, providing an effective noninvasive auxiliary tool for preoperative pathological grading prediction in clinical practice. However, the model's robustness and multi-center applicability need further validation before clinical use."