Deep Learning Identification of Clear Cell Renal Cell Carcinoma on MR Imaging

To investigate the performance of various convolutional neural networks (CNNs) in identifying clear renal cell carcinoma (ccRCC) on MRI and to compare them with radiologists using the clear cell likelihood score (ccLS) algorithm.

A total of 480 CNN models were retrospectively trained using 1 or 3 (fusion) different types of MR images obtained from 310 patients with pathologically confirmed renal masses (cT1, ≤7 cm). MR images included T2-weighted (T2w), T1-weighted opposed-phase/in-phase (T1wOPIP), and corticomedullary phase (CMphase) contrast-enhanced MR acquisitions. To increase the robustness of classification, a 5-fold cross-validation was performed, and the averaged area under curve (AUC) values were compared among the CNN models. The best performing CNN models were compared with radiologists' performance using the ccLS algorithm.

The performance of CNN models in ccRCC classification was significantly influenced by the choice of models, the type of input images, and image size. The best CNN models for the diagnosis of ccRCC using T2w and all 3 types of images (fusion CNN models) achieved an AUC of 0.79 and 0.8, respectively. Combining these models using logistic regression produced a slightly higher AUC value of 0.83. Radiologists using the ccLS achieved an AUC value of 0.85, which was not statistically different than the T2w+fusion model ( P >0.05). Combining the T2w+fusion model and the ccLS achieved an AUC value of 0.86, which was not different than the AUC value using ccLS alone ( P >0.05).

A CNN model integrating three different types of MR images demonstrated performance comparable to that of radiologists in the diagnosis of ccRCC.