Optical Biopsy of Dysplasia in Barrett's Oesophagus Assisted by Artificial Intelligence

Abstract

Optical biopsy in Barrett's oesophagus (BE) using endocytoscopy (EC) could optimize endoscopic screening. However, the identification of dysplasia is challenging due to the complex interpretation of the highly detailed images. Therefore, we assessed whether using artificial intelligence (AI) as second assessor could help gastroenterologists in interpreting endocytoscopic BE images. First, we prospectively videotaped 52 BE patients with EC. Then we trained and tested the AI pm distinct datasets drawn from 83,277 frames, developed an endocytoscopic BE classification system, and designed online training and testing modules. We invited two successive cohorts for these online modules: 10 endoscopists to validate the classification system and 12 gastroenterologists to evaluate AI as second assessor by providing six of them with the option to request AI assistance. Training the endoscopists in the classification system established an improved sensitivity of 90.0% (+32.67%, p < 0.001) and an accuracy of 77.67% (+13.0%, p = 0.020) compared with the baseline. However, these values deteriorated at follow-up (-16.67%, p < 0.001 and -8.0%, p = 0.009). Contrastingly, AI-assisted gastroenterologists maintained high sensitivity and accuracy at follow-up, subsequently outperforming the unassisted gastroenterologists (+20.0%, p = 0.025 and +12.22%, p = 0.05). Thus, best diagnostic scores for the identification of dysplasia emerged through human-machine collaboration between trained gastroenterologists with AI as the second assessor. Therefore, AI could support clinical implementation of optical biopsies through EC.

Keywords: Barrett’s dysplasia; computer-aided diagnosis; endocytoscopy; machine learning; medical training; surveillance.

Figure 1

Schematic overview of the two cohorts participating in the online training and testing modules. (A) The first cohort of 10 endoscopists participated in the online training and testing modules during the pilot phase to validate the classification system. (B) A second cohort of 12 gastroenterologists participated in the same online training and testing modules: six unassisted gastroenterologists and six AI-assisted gastroenterologists. Lastly, a cross-over was performed in which the unassisted gastroenterologists were presented with Test Set 2 again, this time with AI assistance.

Figure 2

In vivo EC procedure and classification. (A) Sequential assessment of the BE performed during endoscopic examination using EC. (B) Examples of EC metaplasia and EC neoplasia with their corresponding histopathology. Examples of EC images showing microvascular features in metaplastic and neoplastic tissue are shown to the right. The EC metaplasia and EC neoplasia images were used to create endocytoscopic BE datasets. BE: Barrett’s oesophagus; CV: crystal violet; EC: endocytoscopy; H&E: haematoxylin and eosin staining; MB: methylene blue; WLE: white light endoscopy.

Figure 3

Overview of the display and the formation of the output of the CNN architecture. (A) After a request, AI assistance was presented as a score on a continuous scale (0–100), which was the average score of all five models in the ensemble. (B) To develop a CNN architecture for the interpretation of endocytoscopic BE images, we trained and tested in distinct image sets of preprocessed images that were selected from the prospectively acquired dataset.

Figure 4

Diagnostic performance of unassisted gastroenterologists and AI-assisted gastroenterologists during the online modules. (A–C) Overview of the gastroenterologists with AI support (N = 6) and without AI support (N = 6) before training (A), after training (B), and at follow-up (C) relative to the overall AI ROC curves for Test Set 1 and Test Set 2, and both test sets combined. (D) Sensitivity, specificity, and accuracy scores of the unassisted gastroenterologists for Test Set 2 before the cross-over and their scores for Test Set 2 with AI support after the cross-over, and the scores of the AI model alone for Test Set 2. * p = 0.020, ** p = 0.024, *** p = 0.039.