03:00 PM Abstract No. 286 Deep learning based on MR imaging for differentiation of low grade and high grade renal cell carcinoma

Abstract

Pre-treatment determination of renal cell carcinoma (RCC) aggressiveness may help guide clinical decision making, informing who may benefit from surveillance versus intervention with ablation or resection. Fuhrman grade has been shown to be a predictor of survival in RCC patients. We aimed to differentiate low grade (Fuhrman I–II) from high grade (Fuhrman III–IV) RCC using routine MR imaging by training a residual convolutional neural network (ResNet). Preoperative MR images (T2-weighted (T2) and T1-post contrast (T1C) sequences) of 300 patients with RCC lesions in a multicenter cohort were acquired. The overall cohort was divided into training, validation, and testing set (70:20:10 split). A ResNet model combining MR images and three clinical variables (age, gender, and tumor size) was built using bagged probabilities and bagged regression to predict tumor grade. In the bagged probabilities version, the T1C, T2, and clinical variables models predict probability of high grade while in the bagged regression version, these models predict Fuhrman grade (float from 0 to 4). In both cases, the output (either probability or float) are bagged into the final classifier. Final model performance was compared with two human experts. Among the 300 patients included, 107 had high Fuhrman grade tumor and 193 had low Fuhrman grade tumor. Our bagged probabilities model achieved a test accuracy of 80.6% (AUC=0.86) with 72.7% sensitivity and 85.0% specificity. Our bagged regression model achieved a test accuracy of 81.8% (AUC=0.75) with 68.4% sensitivity and 88.9% specificity. In comparison, expert 1 achieved accuracy of 60.1% (AUC= 0.63) with 69.8% sensitivity and 55.7% specificity, and expert 2 achieved accuracy of 55.7% (AUC= 0.57) with 62.3% sensitivity and 52.1% specificity. Deep learning can non-invasively predict Fuhrman tumor grade of RCC using conventional MR imaging in a multi-institutional dataset with high accuracy. This non-invasive technique could help clinicians optimize patient management.