Authors
Annika Engel, Mandy Simon, Pedro Guimaraes, Marcin Krawczyk, Oliver Linn, Frank Lammert, Andreas Keller, Markus Casper
Published in
Zeitschrift fur Gastroenterologie. Volume 64. Issue 11. Pages 1751-1759. Epub Nov 20, 2026.
Abstract
Neuronal-network-based methods have increasingly been applied to various tasks in gastrointestinal endoscopy, but prospective data are scarce. Recently, we published two Deep-Learning (DL) classifiers for the detection of gastric precancerous conditions (GPC) and eosinophilic esophagitis (EoE). Our aim was to investigate real-life performance of these tools in an unselected dataset of consecutive patients.
In total, 650 patients undergoing esophagogastroduodenoscopy were recruited. Endoscopists selected up to six images in predefined scope positions (Loc 1-4 proximal stomach; Loc 5-6 esophagus) for evaluation. Finally, 1681 images of 583 patients were analyzed for GPC. For EoE, 936 images of 550 patients were analyzed.
Among patients, 22/583 had GPC, and 7/550 presented with EoE. For the GPC classifier, there were many false positive predictions, especially for Loc2, whereas false negative tool predictions were relatively infrequent. The EoE tool showed fewer false positive results, false negative predictions were very uncommon. For both diseases the tool is better in avoiding false negative results while endoscopists were better in avoiding false positives.
Clinical utility of the GPC classifier is limited because of false predictions. When only using images from a single endoscopic standard position (Loc1), the algorithm performs satisfactorily. The performance of the EoE algorithm is more stable under real-life conditions.
PMID:
43063928
Bibliographic data and abstract were imported from PubMed on 21 Nov 2026.
Read full publication at:
Please sign in
to see all details.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 2
- Comments 0