Endoscopy 2019; 51(04): S6
DOI: 10.1055/s-0039-1681186
ESGE Days 2019 oral presentations
Friday, April 5, 2019 08:30 – 10:30: Artificial intelligence Club A
Georg Thieme Verlag KG Stuttgart · New York

NEAR FOCUS NARROW BAND IMAGING DRIVEN ARTIFICIAL INTELLIGENCE FOR THE DIAGNOSIS OF GASTROESOPHAGEAL REFLUX DISEASE

S Gulati
1   King's College Hospital NHS Foundation Trust, London, United Kingdom
,
J Bernth
2   King's College London, London, United Kingdom
,
J Liao
2   King's College London, London, United Kingdom
,
D Poliyivets
2   King's College London, London, United Kingdom
,
S Chatu
1   King's College Hospital NHS Foundation Trust, London, United Kingdom
,
A Emmanuel
1   King's College Hospital NHS Foundation Trust, London, United Kingdom
,
A Haji
1   King's College Hospital NHS Foundation Trust, London, United Kingdom
,
H Liu
2   King's College London, London, United Kingdom
,
B Hayee
1   King's College Hospital NHS Foundation Trust, London, United Kingdom
› Author Affiliations
Further Information

Publication History

Publication Date:
18 March 2019 (online)

Also available at
 

Aims:

To develop a near focus (NF-NBI) driven artificial intelligence (AI) model for the diagnosis of Gastoesophageal Reflux Disease (GERD).

Methods:

Patients with symptoms of GERD (recorded using the Reflux Disease Questionnaire (RDQ)) were prospectively recruited over 10 months. Upper endoscopy recorded multiple NF-NBI images, video and biopsies of the lower oesophagus. If endoscopy using High-Definition WLE was normal, a pH-recording capsule was placed. Patients were defined according to Lyon criteria; Erosive oesophagitis (EO);non-erosive reflux disease (NERD);functional heartburn (FH).

Two forms of AI were developed and evaluated to automate regions of interest (ROI) and detect IPCLs and morphology: computer vision (CV) and deep convoluted neural network (DCNN) using Resnet50. DCNN was evaluated using training: unseen testing dataset ratios of 50:50 (3872:4280 images) and 75:25 (6484:1668 images). For the purposes of training the AI models, EO and NERD cases were combined as 'GERD'. A novel combined classifier (CC) of both AI methods was evaluated.

Results:

78 consecutive patients were recruited. n = 68 (46 Female, 44.41+/-12.91 years): GERD n = 27 (EO n = 6, NERD n = 21) and FH n = 41 were analysed. The mean IPCL per ROI count was greater in GERD vs. FH: 33.36+/-5.19 vs. 27.9+/-5.72 p = 0.0002 and was used as the primary diagnostic tool. IPCL morphology for GERD vs. FH: length 16.29 vs. 16.98, p = 0.19; width 7.8 vs. 7.8, p = 0.98; red 118.8 vs. 120.6, p = 0.44; green 110.3 vs. 118, p0.006; blue 90.95 vs. 96.81 p = 0.0016.

With CV: mean IPCLs/ROI (threshold 28.4) had sensitivity, specificity, accuracy 85.2, 58.5, 68.2% for GERD.

With DCNN 50:50 these results were 58%, 86% and 76% respectively. DCNN 75:25 produced 67%, 92%, 83% respectively.

CC improved overall specificity (89.1%) and accuracy (78.1%) but not sensitivity (63%).

Conclusions:

AI using NF-NBI is a novel method for the diagnosis of GERD. With increased data, improvements in diagnostic accuracy is achieved further improved using a CC. This model has the potential to provide a reliable safe single-test diagnosis of GERD.