Barrett Esophagus: AI System Reliably Quantifies in Real-Time

Barrett Esophagus
Barrett Esophagus
Investigators evaluated the accuracy of AI assessment of Prague classification and Barrett area quantification from 3D reconstructions of the esophageal surface from 2D endoscopic imaging.

A novel methodology accurately and automatically extracts Prague classification of circumferential length (C) and maximal length (M) scores for the measurement of Barrett epithelium, highlighting new opportunities for risk stratification and therapy response assessments, according to research results published in Gastroenterology.

For the identification of Barrett esophagus, Prague classification is often recommended as a risk stratification tool to determine the interval for surveillance endoscopy; however, the Prague score is minimally quantitative and often subjective due to the operator dependence. While guidelines recommend surveillance intervals based on Barrett esophagus length, Prague classification ignores islands of columnar lined epithelium, which are present in one-third of patients with Barrett esophagus.

A team of investigators for the Translational Gastroenterology Unit at the University of Oxford in the United Kingdom conducted a study to determine the accuracy of assessing the Prague classification and the Barrett area quantification from automatically generated 3-dimensional (3D) reconstruction of the esophageal surface from 2-dimentional endoscopic imaging.

The researchers used a depth estimator network to estimate the endoscope camera distance from gastric folds, segmented the area of Barrett epithelium to measure C&M scores, and tested a derived endoscopy artificial intelligence (AI) system on varying areas of Barrett esophagus and 194 high-definition videos from 131 patients with C&M scores.

The purpose-built 3D-printed esophagus phantom with varying areas of Barrett epithelium provided video data with 97.2% accuracy (marginal ± 0.9 mm average deviation) for C&M measurements and a 98.4% accuracy (± 0.4 cm2 average deviation) for area of Barrett epithelium compared with ground-truth.

Compared with expert scores, the C&M measurements generated by the system were similar, with an overall relative error of 8% (mean difference, 3.6 mm) for C scores and 7% (mean difference, 2.8 mm) for M scores.

“This novel AI system will enable the monitoring of temporal morphological changes of Barrett’s esophagus during development or possible regression in response to treatment,” the authors noted. “Quantification of the Barrett’s area can be used to assess treatment efficacy after ablative treatment of dysplastic Barrett’s esophagus such as radiofrequency ablation, cryoablation, argon plasma coagulation or stepwise endoscopic resection.”

“[The deep learning-based AI system] holds the potential of enhancing endoscopy reporting by providing quantitative and objective data that can be used for review and the assessment of disease progression,” concluded the researchers.

Disclosure: Multiple authors declared affiliations with industry. Please refer to the original article for a full list of disclosures.

Reference

Ali S, Bailey A, Ash S, et al for the TGU Investigators. A pilot study on automatic 3D quantification of Barrett’s esophagus for risk stratification and therapy monitoring. Gastroenterol. Published online June 8, 2021. doi:10.1053/j.gastro.2021.05.059