Construction And Evaluation Of A Deep Learning-based Pathological Differentiation System For Renal Oncocytoma

Objective:Renal Oncocytoma(RO)is the common benign renal tumor,accounting for 3-7% of all renal solid tumors,and its incidence can reach up to 18% for renal tumors less than 4cm.Conservative treatment can be adopted in clinical practice,but it is very difficult to distinguish preoperatively between RO and renal cell carcinoma(RCC),which leads to many unnecessary surgeries,increasing patient suffering and waste of medical resources.The purpose of this study is to explore the diagnosis of RO and RCC based on deep learning technology in enhanced CT images.Methods:A retrospective selection was made of contrast-enhanced CT images from 225 malignant renal tumor and 62 RO patients treated at our hospital from January 2012 to January 2022,with a total of 12,032 images obtained during the cortical,medullary,and excretory phases.The 287 patients were divided into cohort 1 and cohort 2,with cohort 1 forming the training and validation sets and cohort 2 forming a separate external test set.In cohort 1,the training and validation sets were randomly divided in a 4:1 ratio,maintaining a roughly equal proportion of positive and negative labels.We use a deep learning model called Faster R-CNN to build our system,which is composed of the feature extraction network Res Net-50,a region proposal network,and an ROI pooling layer.We import the training dataset into the Faster R-CNN model and train it to recognize and differentiate between RO and RCC enhanced CT images.The predictive ability of the model was evaluated using the test set data,with evaluation indicators including area under the receiver operating characteristic(ROC)curve(AUC),sensitivity,specificity,positive predictive value,negative predictive value,etc.The diagnostic performance of the model on different CT sequences was evaluated and compared with that of human.Results:The fitted model performed well on the test set,with an AUC of 0.804(95% CI,0.768-0.839)for the identification of RO,sensitivity of 0.662,specificity of 0.837,accuracy of0.801,positive predictive value of 0.508,and negative predictive value of 0.907.The AUCs for cortical,medullary,and excretory phase images were 0.894,0.885,and 0.851,respectively.Conclusion:The RO diagnosis model based on deep learning and preoperative enhanced CT can accurately identify RCC and RO,helping and reference for clinical doctors in preoperative decision-making.