Logo image
Back
Mo1249 HIGH ACCURANCY AND EFFECTIVENESS WITH DEEP NEURAL NETWORKS AND ARTIFICIAL INTELLIGENCE IN DETECTION OF EARLY ESOPHAGEAL NEOPLASIA IN BARRETT’S ESOPHAGUS: AN EXTERNAL VIDEO VALIDATION STUDY
Abstract   Open access   Peer reviewed

Mo1249 HIGH ACCURANCY AND EFFECTIVENESS WITH DEEP NEURAL NETWORKS AND ARTIFICIAL INTELLIGENCE IN DETECTION OF EARLY ESOPHAGEAL NEOPLASIA IN BARRETT’S ESOPHAGUS: AN EXTERNAL VIDEO VALIDATION STUDY

Alyssa Y. Choi, Rintaro Hashimoto, Kimberly R. Cavaliere, James Requa, Nabil El Hage Chehade, Andrew Ninh, Tyler Dao, Michael Lugo, Daniel Mai, Kenneth J. Chang, …
Gastrointestinal endoscopy, Vol.91(6 Supplement), pp.AB397-AB397
06/2020
DOI: 10.1016/j.gie.2020.03.2443
url
https://doi.org/10.1016/j.gie.2020.03.2443View
Published (Version of record) Open Access

Abstract

Background The gold standard and most widely used approach for screening and surveillance of Barrett’s esophagus (BE) is esophagogastroduodenoscopy. However, the visual detection of early esophageal neoplasia (high grade dysplasia and T1 stage adenocarcinoma) in BE with white light and virtual chromoendoscopy is still often difficult. We previously designed a convolutional neural artificial intelligence algorithm to detect and localize dysplasia seen on still images. The aim of this study is to assess if this system can detect early esophageal neoplasia in BE in endoscopic video. Methods The AI system, a deep convolutional neural network with base architecture based on Xception with pre-trained weights on ImageNet, was trained on over 4000 unique images from 150 Barrrett’s Esophagus patients. 20 video clips (ranging from 1min to 6 mins) were used for validation from 20 unique patients. 10 patients had atleast 1 dysplastic lesion and 10 patients had non-dysplastic Barretts Esophagus. The videos, containing both white light and narrow band imaging endoscopy, were obtained from an outside institution and therefore were completely unique to the algorithm’s training database. Dysplastic lesions in the video were identified and time stamped by two expert endoscopists and were scored on a scale of subtlety from S1 to S5 (least subtle). The algorithm was designed to conduct continuous real time assessment with generation of atleast 30 predictions per second during video flow. For a prediction to be considered a “clinical prediction” for dysplasia the algorithm had to display a bounding box over a region for > 2 consecutive seconds. To assess false positive (FP) rate and true negative (TN) rate in the non-dysplastic videos, the ratio of FP frames to TN frames were compared as well as the presence of false positive clinical predictions. Results For the dysplastic videos, the algorithm demonstrated 91% per lesion sensitivity in detecting 10/11 lesions. The lesion that was not detected was an S1 lesion. In 2 patients there were two “lesions” identified that were considered false negative clinical predictions. For the non-dysplastic videos, there were 27559 true negative frames and 1045 false positive frames and thus a 3.7% false positive rate. Out of 10 non-dysplastic patients, there were no false positive clinical predictions. Conclusion This external validation video study shows promising results with the AI algorithm demonstrating high sensitivity for dysplastic lesion detection while maintaining a low rate of false positive predictions. Strengths of this system include a true real-time video analysis that does not require freezing endoscopy to generate predictions. The algorithm appears ready for prospective real-time testing.

Details

Metrics

5 Record Views
Logo image