The inability to determine
aggressiveness of RCC based on pretreatment imaging makes it challenging for
physicians to select best benefit treatment. We aimed to
differentiate low grade (Fuhrman I–II) from high grade (Fuhrman III–IV) RCC
using a deep learning model based on routine MR imaging. 297 patients with 300 RCC lesions in a multicenter cohort were included.
A residual convolutional neural network model combining MR images and three
clinical variables was built, which demonstrated high accuracy when compared to
expert evaluation. Deep learning can non-invasively predict Fuhrman grade of
RCC using conventional MR imaging in a multi-institutional dataset.