A weakly supervised deep learning model based on CT images for predicting WHO/ISUP nuclear grade of clear cell renal cell carcinoma

Purpose: The study aims to evaluate the performance of weakly supervised deep learning models in distinguishing high-grade and low-grade clear cell renal cell carcinoma (ccRCC) in comparison to strongly supervised deep learning models. Method: Pathologically confirmed ccRCC from two hospitals between January 2017 and April 2022 were included. The strongly and weakly supervised deep learning models were on three-phase images (CMP, NP and UP) based on the 3D ResNet-18 network and 2D ResNet-18 network, respectively using three-phase images (CMP, NP and UP) to form six deep learning models. Accuracy, sensitivity, specificity, and area-under-the-curve (AUC) were used to assess the discriminatory efficacy of the deep learning models. Results: A total of 306 ccRCC patients were collected in this retrospective study. Among them, 165 were low-grade ccRCC, and 141 were high-grade ccRCC. Data were divided into a training set (n=214) and a testing set (n=92) according to the ratio of 7:3. Among the weakly supervised deep learning models based on three-phase images, the model based on CMP images has the highest diagnostic performance, and its accuracy, sensitivity, specificity, and AUC values are 0.859, 0.857, 0.860 and 0.907, respectively. These values had no significant difference from the strongly supervised deep learning model based on CMP images (the accuracy, sensitivity, specificity, and AUC values were 0.848, 0.854, 0.843 and 0.922, respectively). Conclusion: The weakly supervised deep learning model developed in this study using CMP images has the same high diagnostic performance as the strongly supervised deep learning model in distinguishing high-grade ccRCC from low-grade ccRCC. With richer samples and sufficient developing, the weakly supervised deep learning model may become a routine clinical tool to reduce the physical toll of biopsy on patients.