WCE 2017: Automated and dynamic classification of bladder cancer using deep learning on real-time confocal laser endomicroscopy images

Vancouver, Canada (UroToday.com) Chang et al. present a proof-of-concept study utilizing convolutional neural networks (CNN) to provide automated, intraoperative diagnosis from confocal laser endomicroscopy (CLE) images of the bladder. Their CNN was trained using 170,712 images obtained from 81 patients undergoing a routine cystoscopy and/or a transurethral resection of a bladder tumor. The ground truth was obtained from pathology reports.

The accuracy of their CNN was 87% with a sensitivity and specificity of 79% and 90%, respectively. In comparison, the accuracy of urologist-read CLE images was 79% with a sensitivity and specificity of 77% and 82%, respectively. Although the differences were not statistically significant, these preliminary results show the feasibility of applying CNN to CLE.

These authors are to be commended for their innovative use of deep learning to automate the reading of CLE images. Of course, further studies are needed to compare the use of CNN-assisted CLE with other intraoperative.

Authors: Timothy C. Chang, Darvin Yi, Daniel Rubin, and Joseph C. Liao – Stanford University, Palo Alto, CA, USA

Written by: Michael Owyong (@ohyoungmike), LIFT Fellow, Department of Urology, UC Irvine Medical Center, Orange, CA, USA at the 35th World Congress of Endourology– September 12-16, 2017, Vancouver, Canada.

Login