CT-based Deep Learning for Preoperative Prediction of Pathological Grading of Renal Clear Cell Carcinoma

Abstract

Abstract Objective To construct a model for preoperative noninvasive prediction of WHO/ISUP grading of renal clear cell carcinoma using deep learning technology combined with four-phase CT images of RCCC, and to evaluate the efficacy of the model. Methods CT images, pathological data and general clinical data of 158 patients with renal clear cell carcinoma (RCCC) from January 2019 to June 2021 were collected. Of these 158 patients, 128 were of low grade (WHO/ISUP grade I-II) and 34 were of high grade (WHO/ISUP grade III-IV). The patients recruited were randomly divided into the training set, the validation set and the test set at a ratio of 8:1:1, and CT images were preprocessed first. The ResNet34 network was applied to build a model for predicting WHO/ISUP grading of renal clear cell carcinoma. The validation set data was used for the tuning of model parameters during training, and then the various models built were tested using the test set images; In addition, the accuracy of the test (ACC) and the area under the receiver operating characteristic curve (AUC) were used to evaluate the model performance, and the optimal performing model was screened. Subsequently, the optimal model screened was introduced into the SENet attention mechanism module for model optimization, and the optimized model was retrained and tested; Finally, the ACC and AUC before and after model optimization were compared to further select the optimal performing prediction model. Results The models constructed by the ResNet34 network in the corticomedullary, parenchymal and excretory phase images presented preferable prediction validation, with a prediction accuracy greater than 0.8, while those constructed in the parenchymal phase data demonstrated optimal performance, with a prediction accuracy of 0.867, 0.857 for AUC for patients of low grade, and 0.853 for AUC for patients of high grade. After adding the SE attention mechanism, the optimized SE-ResNet34 model was obtained, and the prediction accuracy of the model improved from 0.867 to 0.878, and that of AUC improved from 0.857 to 0.929 for patients of low grade and from 0.853 to 0.927 for patients of high grade. Conclusion The SE-Resnet34 model based on parenchymal CT boasts a preferable differentiation of WHO/ISUP grade of clear cell renal carcinoma, providing an effective auxiliary means for noninvasive preoperative prediction of pathological grading of renal clear cell carcinoma in clinical practice.