An artificial intelligence (AI)-based algorithm called CADe had a higher accuracy than endoscopists for detecting Barrett neoplasia in endoscopic videos in real time. These findings were published in Gastrointestinal Endoscopy.
The study was conducted at 4 hospitals in the United Kingdom and Europe. The AI model was trained and validated using 3 independent datasets for image and video training, image-based validation, and video-based validation. The training data were collected from 161 patients with Barrett esophagus. The AI model performance in real time was compared with that of 6 endoscopists.
During the image-based validation step, images from 20 patients with neoplastic Barrett and 14 with non-neoplastic Barrett were evaluated. The AI model correctly classified 102 of the 107 neoplastic images and 344 of the 364 non-neoplastic images, corresponding with a diagnostic accuracy of 94.7%, sensitivity of 95.3%, and specificity of 94.5%. For the neoplastic images, the AI model localized all lesions with a recall of 0.71, precision of 0.50, and intersection over union (IoU) of 0.41.
The imaging processing took 5 ms per image to detect lesions and 33 ms per image to localize the lesions.
In the video-based validation step, videos acquired using Fujifilm, Olympus, and Pentax platforms from 32 patients with neoplastic Barrett and 43 with non-neoplastic Barrett were evaluated. The AI model correctly classified 30 of the 32 videos with neoplastic Barrett and 39 of 43 of the videos with non-neoplastic Barrett. The real-time video algorithm had an accuracy of 92.0%, sensitivity of 93.8%, specificity of 90.7%, negative predictive value (NPV) of 95.1%, and area under the curve (AUC) of 0.948.
The endoscopists correctly diagnosed 54 of the 75 videos, corresponding to an accuracy of 71.8%, sensitivity of 63.5%, specificity of 77.9%, and NPV of 74.2%.
Compared with the endoscopists, the AI model had superior accuracy (P <.001), sensitivity (P <.001), specificity (P =.028), and NPV (P =.002).
The AI model was also able to localize lesions in the videos with an accuracy of 83.3%, precision of 0.43, and IoU of 0.41.
In a subgroup analysis of videos from patients with histologically confirmed lesions, the AI model had a superior sensitivity than the endoscopists for classifying low-grade dysplasia (100.0% vs 47.9%; P <.001), high-grade dysplasia (90.9% vs 68.2%; P =.003), and submucosal invasion (100.0% vs 75.0%; P <.001), respectively. The AI algorithm was also able to better detect flat lesions (sensitivity, 92.6% vs 58.0%; P <.001) but not non-flat lesions (sensitivity, 100.0% vs 93.3%; P =.165) compared with the endoscopists, respectively.
This study may have been limited by the fact that the training and validation datasets were collected at a single center.
“This study describes the robust development of a state-of-the-art deep learning model for detection of early Barrett’s neoplasia (CADe) during endoscopy and the results of the first video-based external validation of AI in this area,” the study authors wrote. “Our system achieved high sensitivity, specificity and accuracy and performed significantly better than general endoscopists in a well powered comparison. Our findings demonstrate the feasibility of real-time implementation, as well as the potential added value of AI assisted flat neoplasia detection during endoscopic Barrett’s surveillance.”
Disclosure: Some study authors declared affiliations with biotech, pharmaceutical, and/or device companies. Please see the original reference for a full list of authors’ disclosures.
Abdelrahim M, Saiko M, Maeda N, et al. Development and validation of artificial neural networks model for detection of Barrett’s neoplasia, a multicenter pragmatic non-randomized trial. Gastrointest Endosc. Published online October 22, 2022. doi:10.1016/j.gie.2022.10.031