O2 Staging of Barrett’s neoplasia using artificial neural networks, proof of concept study

Abstract

<h3>Introduction</h3> Endoscopic differentiation between intramucosal and submucosal Barrett’s neoplasia has several important implications but remains challenging even for expert endoscopists. Recent studies demonstrated promising results on AI-assisted detection of Barrett’s neoplasia, but data on AI-assisted staging is limited. We aimed to develop and validate an AI system for classification of Barrett’s neoplasia into intramucosal or submucosal, and compare its performance to expert endoscopists. <h3>Methods</h3> The model, based on VGG-16 architecture, was trained on 117 images of prospectively collected and annotated Barrett’s neoplastic lesions. Rotation and random flip were used for data augmentation. The ground truth was the histological staging of endoscopically resected specimens performed by two pathologists with expertise in Barrett’s neoplasia. Images comprised of WLI, enhanced imaging, and magnification views. The model was designed to classify images as either intramucosal (pT1a) or submucosal (pT1b). Performance of the AI system was compared to a group of three experts. <h3>Results</h3> The AI model was tested on an independent dataset of 90 images. The accuracy, sensitivity and specificity and AUC of the AI model in differentiating between intramucosal and submucosal neoplasia was 70.9%, 72.5%, 65.7%, and 0.781 respectively. Mean accuracy, sensitivity and specificity of experts were 73.3%, 63.3% and 83.3% respectively. Processing speed of the AI system was 5 ms/image. <h3>Conclusion</h3> This study demonstrates the feasibility of AI-assisted staging of Barrett’s neoplasia on endoscopic images. The AI model’s performance was comparable to that of experts. More work is needed to further develop this early model and validate its use on real-time video sequences.