To build a convolutional neural network model to predict tumor necrosis on whole-slide images (WSI) in clear cell renal cell carcinoma, and to assess the performance of the model.

25 WSIs of clear cell renal cell carcinoma with tumor necrosis are downloaded from TCGA database, and 100 pathological slides of clear cell renal cell carcinoma from the First Affiliated Hospital of Sun Yat-sen University are scanned. Tumor and necrosis region are annotated in WSIs under the guidance of pathologists. The WSIs from TCGA are then cropped into 21 411 patches (2 556 necrosis and 18 855 non-necrosis), training the convolutional neural network DenseNet201 to establish a deep learning model for detecting tumor and necrotic patches. 3 855 patches from our center are used for external testing. Based on the classification results and location information of patches, tumor necrosis-prediction maps are generated. The quantile threshold method is used to analyze the performance of the model to detect tumor necrosis at the whole slide level.

In the external testing group, the accuracy, sensitivity, specificity and AUC of the model at patch-wise are 0.909, 0.890, 0.943, and 0.973 for tumor and non-tumor classification. The accuracy, sensitivity, specificity and AUC are 0.717, 0.958, 0.657 and 0.920 respectively for necrotic and non-necrotic classification. The AUC, sensitivity, specificity and accuracy of tumor necrosis at slide-wise are 0.903, 0.760, 0.933, and 0.890. Based on the classification results of patches, prediction maps of tumor necrosis are generated and the percentage of tumor necrosis is calculated.

The classification model based on convolutional neural network algorithm can be used as an assistant diagnostic tool to help pathologists detect tumor necrosis clear cell renal cell carcinoma and quantitatively evaluate the extent of necrosis.