Prediction Of Pathological Grading And Molecular Typing Of Clear Cell Renal Cell Carcinoma Using Deep Learning Based On CT Images

Background Renal cell carcinoma(RCC)accounts for about 3%of all adult malignant tumors,and the incidence of renal cell carcinoma increases gradually with the development of social economy.There are many subtypes of RCC,and clear cell renal cell carcinoma(ccRCC)is the most common subtype with notorious heterogeneity.It is necessary to carry out histopathological grading and molecular typing,which can assist in the formulation of diagnosis and treatment plan and predict prognosis.Surgical resection of focal ccRCC is the first choice.Partial nephrectomy can improve the overall survival and reduce the risk of surgery compared with radical nephrectomy.Therefore,clinical evaluation of renal tumor morphology and heterogeneity is needed.The identification ability of Deep learning is better than humanbeings.Deep learning uses quantized image information to identify complex patterns,and has recently flourished in the application of medical imaging.Such as differentiating between benign and malignant kidney mass,the accuracy of deep learning models even surpasses four senior radiologists.Deep learning techniques represented by deep convolutional neural networks show great potential in assisting clinical decision making.Objuctive To reduce the manual segmentation load and achieve a relatively generalized deep learning segmentation algorithm,this study first develop a semantic segmentation model for renal tumors based on CT images and deep learning,and then uses the automatic segmentation and manual re-examination to achieve accurate segmentation.The prediction models of ccRCC pathological grading and molecular subtypes are constructed using segmentation results to provide auxiliary tools for clinical diagnosis and treatment.Materials and Methods The CT images of 190 renal tumors with artificial labeling were collected by KiTS19 open data set.Preoperative CT images and clinicopathological information from 211 patients with WHO/ISUP grading were collected retrospectively from two hospitals.140 cases containing DNA methylation subset were collected from TCGA-KIRC data collection.The KiTS19 data set was used to establish the renal tumor segmentation model based on 3D-Unet and 3D-Res-Unet network architecture,and the evaluation indices were compared.The 3D-Res-Unet model was fine-tuned by manually tagging cases from local data.The remanent cases is automatically segmented by fine-tuned segmentation model,and then manually checked one by one.Based on the fine segmented image and the 3D-Resnet-18 network,the ccRCC pathological grading and molecular typing models were constructed respectively.The diagnostic efficacy of these models were evaluated by receiver operating characteristic curve.Results The mean dice similarity coefficient(DSC)of renal tumor segmentation models based on 3D-UNet and 3D-Res-UNet were 0.759 and 0.814(P<0.05),respectively..The mean DSC of renal tumor segmentation models pre-fine-tuning and post-fine-tuning for the local clinical test set were 0.718 and 0.835 respectively.The area under curv(AUC)of ccRCC pathological grading model and molecular typing model evaluated using the test set were 0.902 and 0.886(with 95%confidence intervals of 0.781-1.000 and 0.715-0.985)respectively.Conclusion This study demonstrates that the segmentation of kidney and kidney tumor using 3D-Res-UNet architecture is feasible.Using the open data set to establish the segmentation model,and then adding the local clinical data set to fine-tune the model is a feasible way to reduce the labor cost and build a more generalized model.The algorithm model based on deep learning can precisely predict ccRCC WHO/ISUP grading and DNA methylation typing and evaluate the overall tumor more comprehensively.Finally helping guide clinical diagnosis and treatment and prognosis prediction.