A deep-learning based artificial intelligence (AI) approach for differentiation of clear cell renal cell carcinoma from oncocytoma on multi-phasic MRI

Abstract

Purpose: To investigate the diagnostic performance of a deep convolutional neural network for differentiation of clear cell renal cell carcinoma (ccRCC) from renal oncocytoma.

Methods: In this retrospective study, 74 patients (49 male, mean age 59.3) with 243 renal masses (203 ccRCC and 40 oncocytoma) that had undergone MR imaging 6 months prior to pathologic confirmation of the lesions were included. Segmentation using seed placement and bounding box selection was used to extract the lesion patches from T2-WI, and T1-WI pre-contrast, post-contrast arterial and venous phases. Then, a deep convolutional neural network (AlexNet) was fine-tuned to distinguish the ccRCC from oncocytoma. Five-fold cross validation was used to evaluate the AI algorithm performance. A subset of 80 lesions (40 ccRCC, 40 oncocytoma) were randomly selected to be classified by two radiologists and their performance was compared to the AI algorithm. Intra-class correlation coefficient was calculated using the Shrout-Fleiss method.

Results: Overall accuracy of the AI system was 91% for differentiation of ccRCC from oncocytoma with an area under the curve of 0.9. For the observer study on 80 randomly selected lesions, there was moderate agreement between the two radiologists and AI algorithm. In the comparison sub-dataset, classification accuracies were 81%, 78%, and 70% for AI, radiologist 1, and radiologist 2, respectively.

Conclusion: The developed AI system in this study showed high diagnostic performance in differentiation of ccRCC versus oncocytoma on multi-phasic MRIs.

Keywords: Clear cell renal cell carcinoma; Deep learning; Multi-phasic MRI; Oncocytoma; Radiomics.

Fig. 1.

Study flow chart summarizing the patient and lesion selection.

Fig. 2.

Showing segmentation process and bounding box generated of the tumor shown in Fig. 2. Tumor was located on T2 WI (A), a seed point was placed on the lesion based on rough estimation of its location (B). Bounding box was generated, relative to the image (C), and extracted (D).

Fig. 3.

Convolutional neural network architecture.

Fig. 4.

Receiver operating characteristic (ROC) curve plotted for performance of algorithm (blue for full data and orange for 80 lesion subset) for differentiating ccRCC from oncocytoma. Gray and yellow lines show radiologist performance on 80 lesion subset. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 5.

(A) T2-WI, (B) pre-contrast phase, (C) post-contrast arterial, and (D) post-contrast venous phase MRI image sets in a 34-year old male with a clear cell renal cell carcinoma diagnosed with von Hippel-Lindau syndrome. The mass was correctly identified by both radiologists and AI with high probability of being ccRCC.

Fig. 6.

(A) T2 WI, (B) pre contrast phase, (C) post-contrast arterial, and (D) post contrast venous phase MRI image sets in a 41 year-old male with an oncocytoma tumor diagnosed with Birt-Hogg-Dubé syndrome. This mass was misclassified by both radiologists as ccRCC, but was correctly diagnosed by the AI algorithm. Imaging features of the mass represent common enhancement patterns similar to that of ccRCC.