Description

Background and Rationale

Third-space endoscopy procedures are technically demanding. The primary challenge lies in the inadvertent transection of submucosal vessels, which leads to bleeding that obscures the surgical field and increases the risk of perforation. Currently, vessel identification is entirely operator-dependent.

Our team has developed a deep-learning based artificial intelligence model trained on 250,000 annotated images from 150 POEM procedures. This model is optimized for minimal latency, allowing for real-time visual overlays (delineation) of blood vessels on the endoscopic monitor.

Study Objectives The primary objective is to evaluate the Vessel Detection Rate (VDR)-the proportion of vessels identified by the endoscopist when assisted by the AI compared to standard practice.

The study will also investigate:

Vessel Detection Time (VDT): The latency between a vessel appearing in the field of view and its identification.

Study Design \& Workflow:

In this prospective study, the AI system will be integrated into the Olympus EVIS X1 series endoscopy stack. As the endoscopist navigates the submucosal space, the AI will provide real-time visual segmentation masks highlighting vessels. The performance will be recorded and compared against a post-procedure review by independent experts to calculate sensitivity and detection speed.