Article Text

Download PDFPDF
Real-time unblinding for validation of a new CADe tool for colorectal polyp detection
  1. Pieter Sinonquel1,2,
  2. Tom Eelbode3,
  3. Cesare Hassan4,
  4. Giulio Antonelli4,
  5. Federica Filosofi4,
  6. Helmut Neumann5,
  7. Ingrid Demedts1,2,
  8. Philip Roelandt1,2,
  9. Frederik Maes3,
  10. Raf Bisschops1,2
  1. 1 Department of Gastroenterology and Hepatology, University Hospitals Leuven, Leuven, Belgium
  2. 2 Department of Translational Research in Gastrointestinal Diseases (TARGID), KU Leuven, Leuven, Belgium
  3. 3 Department of Electrical Engineering (ESAT/PSI), Medical Imaging Research Center, KU Leuven, Leuven, Belgium
  4. 4 Digestive Endoscopy Unit, Nuovo Regina Margherita Hospital, Rome, Italy
  5. 5 Department of Gastroenterology and Hepatology, University Medical Center Mainz, Mainz, Germany
  1. Correspondence to Professor Raf Bisschops, Gastroenterology and Hepatology, KU Leuven University Hospitals Leuven, Leuven 3000, Belgium; raf.bisschops{at}

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.

Summary box

What is already known about this subject?

  • CADe has been claimed to marginalise polyp miss rate due to failed polyp recognition due to its high accuracy in identifying and localising a polyp present on an endoscopic frame. CADe validation has been mainly performed in an artificial setting by standalone performance against expert endoscopists. The main limitation of this methodology is not only the lack of human–machine interaction as in a clinical setting, but also the fact that the ground truth is represented by polyps recognised only by humans, preventing polyps missed by the endoscopists but CADe detected to become reference standard.

What are the new findings?

  • We validated the diagnostic performance of a new CADe system that is novel and innovative in its design since it has a temporal algorithm integrated, resulting in a ‘memory of previous processed frames’ leading to a lower false-positive rate and is developed by using a propagation tool to increase the number of allocated images. In addition, we introduce a novel study design with real-time unblinding. By using a composite reference standard represented by both human and CADe detection, we showed in real-time a very high clinical accuracy of 96.5% of our CADe system for polyp detection. In a challenging setting with real-time endoscopy, our system was not inferior to endoscopists with a high adenoma detection rate (>35%). Unexpectedly, we found that the miss rate due to failure of polyp recognition in high detectors is in fact very low (<2%) and lower than expected from the literature (20%–25%).

How might it impact on clinical practice in the foreseeable future?

  • The confirmation of clinical accuracy for polyp detection is reassuring to use this new technological background for CADe development and may set a new standard. Our study design with real-time unblinding allows to set a new gold standard for testing the real additional clinical value of CADe over human detection in a real-life clinical setting.


In this prospective multicentre …

View Full Text


  • Correction notice This article has been corrected since it published Online First. The seventh author's name has been corrected.

  • Contributors PS, RB, TE and CH are responsible for the concept of this paper and wrote the manuscript. PS, RB, CH, HN and TE developed the study protocol. TE and FM developed the CADe tool. PS, CH, GA, FF, HN and RB enrolled patients. CH, GA, HN, DI, PR and RB performed endoscopic procedures. PS, TE, GA and FF checked the CADe screen. All authors provided valuable feedback, suggestions and corrections to improve the quality of the manuscript. The manuscript is approved by all authors.

  • Funding RB and TE are supported by a grant of Research Foundation Flanders. PR is supported by Clinical Mandate from Belgian Foundation against Cancer (Stichting tegen Kanker). FM is supported by KU Leuven internal grant C24/18/047 and by the Flemish Government (AI Research Program).

  • Competing interests PS is supported by an unrestricted research grant from Pentax. RB received speaker’s fees, consultancy and research support from Pentax, Fujifilm and Medtronic. CH and HN received consultancy support from Fujifilm and Medtronic.

  • Patient and public involvement Patients and/or the public were not involved in the design, or conduct, or reporting, or dissemination plans of this research.

  • Patient consent for publication Not required.

  • Provenance and peer review Not commissioned; internally peer reviewed.