2. UroToday Home
  3. Conference Highlights
  4. EAU 2018
  5. EAU 2018 Bladder Cancer

EAU 2018

EAU 2018: Computer-Assisted Diagnosis During Blue Light Cystoscopy Using Image Analysis Methods: Ahead of Pathology?

Copenhagen, Denmark (UroToday.com) Dr. Kriegmair and colleagues discussed computer assisted diagnosis during blue light cystoscopy. Blue light cystoscopy increases the detection of bladder tumors, in particular for small and flat CIS lesions, however the drawback is a rather low specificity. Furthermore, surgeons cannot reliably distinguish between tumor stage and grading during the intervention. As such, the aim of the study was to assess the possibility of a computer-aided diagnosis of suspicious lesions during blue-light cystoscopy using image analysis and machine learning approaches.

For this study, 122 digital endoscopic images of blue light cystoscopy showing suspicious lesions of the urinary bladder were collected from three urological hospitals. Standard pathology reports were assessed for final diagnosis. For the automatic analysis of endoscopic images a data pipeline was constructed combing data pre-treatment, image feature extraction and supervised classification methods. As a classifier, a combination of principal component analysis with linear discriminant analysis and a combination of principal component analysis and support vector machine were tested.

Routine pathology revealed transitional carcinoma of the bladder in 86 cases, which included 48 low grade tumors, 29 high-grade tumors and 9 CIS cases. Additionally, 16 pictures showed inflammatory tissue and 20 normal urothelium. The linear discriminant value 1 was significantly different between high and low-grade tumors (-0.74±1.55 vs. 1.01±1.59, p<0.0001). Prediction of a high-grade pathology using the principal component analysis/linear discriminant analysis model showed a sensitivity and specificity of 72.4% and 87.5%. To distinguish between inflammation and CIS Dr. Kreigmair used a principal component analysis/ support vector machine model, which revealed a sensitivity and specificity of 88.9% and 81.3% (predictive accuracy of 84.0%).

The authors concluded that the use of image analyses and machine learning approaches enables differentiation between high-grade and low-grade urothelial carcinoma of the bladder from endoscopic images of blue-light cystoscopy. It may also allow a better discrimination of flat lesions. Future real-time, intraoperative employment of this technology might increase specificity of blue light cystoscopy and help to stratify the extent of transurethral resection and early intravesical instillation of chemotherapy.


Presented by:  Maximilian C. Kriegmair, MD University Medical Center Mannheim, Mannheim, Germany

Co-Authors: Hartmann A, Todenhöfer T, Ali N, Hipp G, Knoll T, Honeck P, Oberneder R, Stenzl A, Popp J, Bocklitz T

Written by: Zachary Klaassen, MD, Urologic Oncology Fellow, University of Toronto, Princess Margaret Cancer Centre, twitter: @zklaassen_md at the 2018 European Association of Urology Meeting EAU18, 16-20 March, 2018 Copenhagen, Denmark