With increasing discovery of renal incidentalomas, it is important to make a pretreatment differentiation between benign renal masses and malignant tumors. We aimed to develop a deep learning model that distinguishes benign renal tumors from renal cell carcinoma (RCC) by applying a residual convolutional neural network (ResNet) on routine magnetic resonance imaging (MRI). Preoperative MR images (T2-weighted and T1-post contrast sequences) of 1162 renal lesions in a multicenter cohort with definitive pathology were divided into training, validation, and test sets (70:20:10 split). Two ResNet models were trained to predict renal tumor pathology, one on T2-weighted images and one on T1-post contrast images. A logistic regression model was trained based on clinical features of each patient. A final ensemble model based on a bagging classifier was trained on the outputs of the ResNet models and the clinical features model to predict renal tumor pathology. Final model performance was compared with the interpretation of four experts. Among the 1162 renal lesions, 655 were malignant and 507 were benign. The final ensemble model achieved a test accuracy of 0.68 (95% CI: 0.59-0.76), sensitivity of 0.89 (95% CI: 0.79-0.95), and specificity of 0.41 (95% CI: 0.28-0.55). Compared to all experts averaged, the ensemble deep learning model had higher test accuracy (0.68 vs. 0.6, P = 0.12), higher test sensitivity (0.89 vs. 0.8, P = 0.11) and higher test specificity (0.41 vs. 0.35, P = 0.45), although none of these was statistically significant. Deep learning can noninvasively distinguish benign renal tumors from RCC using conventional MRI in a multi-institutional dataset with high accuracy and sensitivity compared with experts.
Read full abstract