Identification of Dysplasia in the Barrett’S Esophagus Using an Endocytoscopy Classification System: Preliminary Results of a Prospective Comparison Between Clinicians and Artificial Intelligence

Aims To investigate the feasibility of endocytoscopy (EC) in differentiating dysplastic from non-dysplastic tissue in the Barrett’s esophagus (BE) in vivo, performance of clinicians and a computer-aided diagnosis (CADx) algorithm were assessed and compared with each other. Ultimately, the potential of the implementation of CADx will be determined during a test in which clinicians can use the help of the CADx.

Methods We performed endocytoscopy prospectively in BE patients; areas of interest were videotaped using EC and included for analysis if frames were classifiable and could be correlated to histology of targeted biopsies. An EC classification system for the BE was developed that differentiated BE metaplasia from BE neoplasia. Online training and examination modules were designed for clinicians (5 BE experts, 5 gastroenterologists and 5 residents) and scored. Finally, a convolutional neural network was trained on the collected frames and tested on a separate test set.

Results Endocytoscopy was performed in 52 BE patients. We selected 728 metaplastic and 824 dysplastic images from 20 patients for training of CADx. Accuracy, sensitivity and specificity of clinicians before training (n = 14) were 62.6 %, 56.2 % and 70.6 % and after training (n = 14) were 78.6 % (P < 0.05), 86.2 % (P < 0.05) and 69.5 % (P > 0.05) respectively. After an interval of at least two weeks, their (n = 9) accuracy and sensitivity significantly decreased to 73.0 % and 75.6 %. The average accuracy, sensitivity and specificity of the algorithm on image basis over 5 runs were 79.6 %, 85.3 % and 74.0 %, respectively.

Conclusions EC allows in vivo discrimination of metaplastic and dysplastic BE tissue. Interpretation is however not straightforward for clinicians and requires training and maintenance. AI shows promising performance in analyzing EC images and can enable highly accurate diagnosis. This could help facilitate generalization of EC in clinical practice.

Citation: van der Laan JJH, van der Putten JA, Zhao X et al. OP178 IDENTIFICATION OF DYSPLASIA IN THE BARRETT’S ESOPHAGUS USING AN ENDOCYTOSCOPY CLASSIFICATION SYSTEM: PRELIMINARY RESULTS OF A PROSPECTIVE COMPARISON BETWEEN CLINICIANS AND ARTIFICIAL INTELLIGENCE. Endoscopy 2021; 53: S73.

Publication History

Article published online:
19 March 2021