AUA 2019: Automated Cystoscopic Detection of Bladder Cancer Using Deep-learning

Chicago, IL ( The gold standard for evaluation and diagnosis of new bladder cancer and for surveillance of patients with non-muscle invasive bladder cancer (NMIBC) remains cystoscopy. However, as has been clearly demonstrated before, cystoscopy is not a perfect test – besides being user dependent, white light cystoscopy has significant limitations. It may miss small tumors, CIS and other non-obvious lesions in certain anatomic locations. Tools to augment white light cystoscopy have been increasing, and use of technology such as blue light cystoscopy (BLC) have been important updates.

In this interesting study, the authors assess the ability to overcome user error and missed disease by utilizing machine learning – pushing towards a more “automated cystoscopy.” They aimed to develop a deep-learning algorithm for augmented detection of bladder cancer during standard cystoscopy. To that end, following IRB approval, the prospectively identified and collected videos of office-based cystoscopy and transurethral resection of bladder tumor from 100 subjects (for a total of 141 videos); each of these videos were then annotated by a reader. To help develop the algorithm, video frames (n=611) containing histologically confirmed papillary bladder cancer were selected and tumor outlined (green line, figure below). Bladder neck, ureteral orifices, and air bubble were labeled for exclusion learning. 

The specific technology utilized was “TUMNet”, an image analysis platform based on convolutional neural networks, that evaluated videos in two stages: 1) recognition of frames containing abnormal areas and 2) segmentation of regions within the frame occupied by tumor. 

They first developed the algorithm using a training set was constructed based on 95 subjects (417 cancer and 2,335 normal frames). A validation set was constructed based on 5 subjects (211 cancer, 1,002 normal frames). 

Following development in the training set, they assessed the algorithm in the validation set of 5 patients. TUMNet per-frame sensitivity was 88% (186/211) and per-tumor sensitivity was 90% (9/10); per-frame specificity of 99% (992/1002). Based on this, the authors were confident in the algorithm.

As a result, they note that an ongoing prospective study is underway to evaluate TUMNet performance. So far, it has accurately detected all 16 tumors that were resected in the ongoing prospective test cohort (n=9; 15 cancer, 1 benign).

Importantly, the authors note that, especially in centers that cannot afford the additional technology (such as BLC, etc), such an algorithm may be useful to augment physician cystoscopies. 

Presented by: Eugene Shkolyar, MD, Stanford, CA

Co-Authors: Xiao Jia, Lei Xing, Joseph Liao

Written by: Thenappan Chandrasekar, MD, Clinical Instructor, Thomas Jefferson University, @tchandra_uromd, @JEFFUrology at American Urological Association's 2019 Annual Meeting (AUA 2019), May 3 – 6, 2019 in Chicago, Illinois