Artificial Intelligence (AI) Vs Endoscopists in Detection of Barrett’s Neoplasia

Aims We aimed to develop and validate an AI algorithm based on deep neural networks for detection of Barrett’s neoplasia, and compare its performance to endoscopists.

Methods The AI algorithm, based on VGG16 architecture, was trained and validated on 65,545 images (96 videos) of neoplastic Barrett’s and 101,5342 images (65 videos) of non-neoplastic Barrett’s. Ground truth was histological diagnosis and expert review. The algorithm was trained to detect and classify images and videos as neoplastic or non-neoplastic. For testing, sample size was calculated using extended McNemar’s test at 90 % power and 5 % significance level, assuming 91 % sensitivity of the AI system (based on validation dataset and pilot study) and endoscopists sensitivity of 68 %. Primary end point was sensitivity of AI diagnosis of Barrett’s neoplasia. We asked 6 endoscopists who regularly perform endoscopic surveillance and therapy of Barrett’s neoplasia to review same videos and classify them into neoplastic or non-neoplastic. We collected and compared metrics on processing speed, sensitivity, specificity, NPV and accuracy.

Tab. (1): Performance of deep learning algorithm in detection of Barrett’s neoplasia compared to endoscopists (n = 75).

	Sensitivity	Specificity	NPV	Accuracy
AI	96.88 %	90.70 %	97.50 %	93.33 %
Endoscopists (n = 6)	72.95 %	83.89 %	78.29 %	78.47 %
P value	P < 0.0001	P < 0.0001	P < 0.0001	P < 0.0001

Results We included 75 (32 neoplastic and 43 non-neoplastic) Barrett’s videos.In the neoplastic videos, 27 (84.3 %) were flat lla/b lesions.The AI system diagnosed Barrett’s neoplasia with sensitivity, specificity, NPV and accuracy of 96.88 %, 90.70 %, 97.50 % and 93.33 % respectively. The average sensitivity, specificity, NPV and accuracy of endoscopists were 72.95 %, 83.89 %, 78.29 % and 78.47 % respectively. AI system’s sensitivity, specificity, NPV and accuracy were significantly better than endoscopists (P < 0.0001). Processing speed of the AI system was 5ms/image.Table (1) summarizes the results.

Conclusions Our data demonstrates the feasibility of AI-based neoplasia detection during Barrett’s assessment. AI was better than endoscopists in detection of Barrett’s neoplasia on recorded videos. The NPV of AI (97.5 %) is very close to the 98 % target set by PIVI. This needs to be validated during real time endoscopic assessment.

Citation: Abdelrahim M, Saikou M, Maeda N et al. OP11 ARTIFICIAL INTELLIGENCE (AI) VS ENDOSCOPISTS IN DETECTION OF BARRETT’S NEOPLASIA. Endoscopy 2021; 53: S9.

Publikationsverlauf

Artikel online veröffentlicht:
19. März 2021