Deep-learning system accurately detects neoplasia in patients with Barrett's esophagus

Last Updated: 2019-12-03

By Will Boggs MD

NEW YORK (Reuters Health) - A new deep-learning computer-aided detection (CAD) system detects neoplasia in patients with Barrett's esophagus with greater accuracy than general endoscopists, researchers report.

"The endoscopic detection of early Barrett's neoplasia is challenging," said Dr. Jacques J. Bergman of Amsterdam UMC at the University of Amsterdam in the Netherlands.

"Early lesions often show only subtle endoscopic abnormalities, and it is difficult to learn (to identify them) since most endoscopists rarely encounter early lesions," he told Reuters Health by email. "If the CAD system were implemented during surveillance endoscopies, it would assist the endoscopist to detect early neoplastic lesions, which would contribute to early treatment and better outcome for the patient."

Dr. Bergman and colleagues developed a deep-learning CAD system for endoscopic images of Barrett's esophagus, validated the system and compared CAD performance with the performance of a group of international endoscopist assessors.

They used five data sets that included nearly 500,000 images overall for pre-training, training, internal validation and two external validations of the deep-learning system.

Internal validation demonstrated 88% accuracy, 88% sensitivity, and 89% specificity for classifying images as early neoplasia versus non-dysplastic Barrett's esophagus.

In two external validations, the system achieved 88-89% accuracy, 90-93% sensitivity and 83-88% specificity for classifying images, the researchers report in Gastroenterology, online November 21.

The final data set was evaluated by the system and by 53 international endoscopist assessors. The general endoscopists achieved 73% classification accuracy, 72% sensitivity, and 74% specificity, which were substantially inferior to the deep-learning CAD system, regardless of the level of endoscopist experience.

The system analyzed images in a mean 0.124 second, while the mean time per image for assessors was 46.3 seconds.

"In the near future, the CAD system will be able to assist the endoscopist to detect early neoplastic lesions and hopefully will contribute to a better outcome for the patient," Dr. Bergman said.

SOURCE: https://bit.ly/2DKyzV2

Gastroenterology 2019.