Can artificial intelligence help to reduce intraprocedural bleedings during third-space endoscopy?

Artificial intelligence (AI) is gradually finding its way into endoscopy practice. This introduction, however, has not progressed as fast as one might expect. The road from development through implementation into clinical practice is long and laborious. Furthermore, important aspects such as scientific proof, compatibility, trust, and legal and financial issues are involved.

Besides the detection, delineation, and characterization of early neoplasms of the gastrointestinal tract, and algorithms measuring and supporting quality aspects of endoscopy, AI algorithms designed specifically for supporting therapeutic third-space endoscopy such as endoscopic submucosal dissection (ESD) and peroral endoscopic myotomy have also been developed. In this issue of Endoscopy, Scheppach et al. describe the preclinical validation of an AI-based decision support algorithm, which is capable of detecting submucosal vessels during third-space therapeutic endoscopy [1]. The technical background and development of the algorithm have been published recently [2].

How would this novel algorithm perform in clinical practice? Would it significantly influence the designated targets of reduction in intraprocedural bleeding and procedure time?

The algorithm was tested under preclinical conditions involving analysis of vessel-enriched videos. Videos simulate clinical conditions better than still images and are now becoming standard in the development and testing of diagnostic algorithms in endoscopy. Videos are much more representative of real-world endoscopy. Under these circumstances the vessel detection rate was found to be significantly higher, and vessels were detected faster than when no AI was available during video review. Another positive outcome was that the performance of endoscopists with little experience in submucosal dissection, as well as experienced endoscopists, improved in terms of vessel detection and speed of vessel detection.

What clinical benefits does this vessel detection algorithm offer? Detecting submucosal vessels early and having the opportunity to apply selective coagulation potentially results in less intraprocedural bleeding, thereby saving time, reducing the risk of coagulation defects, and reducing overall complication risk. Updating the system to incorporate characterization of blood vessel types – veins and arteries – as the authors suggest, could further increase clinical applicability.

AI may support endoscopists, especially those with limited experience or skill. Lack of experience or knowledge of specific anatomy may cease to be an issue, and other factors such as gaze pattern, concentration, self-confidence, and fatigue during relatively long endoscopic procedures may no longer lead to inconsistent outcomes as AI reduces performance variability.

Is vessel recognition a complex task? Recognizing blood vessels during endoscopic resection is not that difficult. The authors based their study design on a vessel detection rate of 75% for experts and 65% for nonexperts, indicating room for improvement; the algorithm previously detected 10% more vessels than experts. The authors found that overall performance improved with the use of AI, from 56% to 72%, demonstrating a significant increase in the vessel detection rate. Timely vessel detection is definitively important for prevention of intraprocedural bleeding and the authors also found that vessel detection was faster with AI support than without it. If bleeding occurs in the submucosal space, the endoscopic view of the resection plane is blurred. Furthermore, detection and treatment of intraprocedural bleeding, including the need for instrument change to coagulation forceps, prolongs the procedure time, while coagulation applied too frequently or intensely during hemostatic treatment can cause damage to the muscle layer, introducing the risk of (late) perforation [3].

Prevention of bleeding during ESD involves endoscope tip control, visualization, and concentration of the endoscopist. The injected solution is important, preferably not too dark in color and with the addition of adrenaline. Careful preparation is essential, as is ensuring that the plane of dissection is correct. Owing to the anatomy of the submucosal vessels, endoscopists should be aware that vessels are less prevalent closer to the muscularis propria layer.

A positive safety contribution of the novel algorithm described by Scheppach et al. is the red border tracking on the screen, which appears when the knife comes into the vicinity of a submucosal vessel. This will definitively alert the endoscopist; however, in practice, the use of and trust in technologies such as these are often threatened by false alarms. In the Scheppach et al. study, false alarms were not prevalent and had a short duration, which could positively influence practicality and uptake of the system.

Limitations of this novel contribution are clearly stated by the authors, including the fact that the system is currently limited to a single endoscope series. As high-definition white-light imagery from the most used brands varies, the algorithm will need separate training and validation for other brands before its use can be expanded into routine practice. Given that only a minority of endoscopists currently perform third-space endoscopy, this is still a niche market.

A remarkable fact is that in 25% of AI-labeled screens, the endoscopist did not recognize the vessel. Clarity of the image or trust in AI have been suggested as potential factors. Nevertheless, the algorithm still complies with the generalized criteria of the European Society of Gastrointestinal Endoscopy of being at least as good as an experienced endoscopist [4].

It is impossible to perform interventional endoscopy without some level of bleeding. From my personal experience during ESD, most bleedings occur during the blind circumferential incision, a stage not covered by this novel algorithm. Besides bleeding, perforation of the muscle layer is still the biggest complication of third-space endoscopy, and it is likely that an algorithm designed to highlight the muscle layer or present a muscle warning could improve safety even further.

Other developments designed to highlight vessels and provide support in prevention and/or treatment of intraprocedural bleeding have the potential to influence clinical practice. Recently, red dichromatic imaging (RDI) and amber-red color imaging (ACI) have been introduced into clinical practice. These are image-enhancing technologies highlighting blood vessels. RDI (Olympus, Tokyo, Japan) utilizes lights of longer wavelengths, which have weak light-scattering characteristics, enhancing blood vessels and making the source of bleeding visible [5]. ACI has been developed by Fujifilm (Tokyo, Japan), and involves the use of amber-red, green, and blue light-emitting diodes (LEDs). The spectral profile of the LED light source is controlled to make differences in the shading of blood easier to see [6]. It will be interesting to see how all these new technologies of image enhancement and AI support will blend into endoscopy practice.

Some questions remain. How would this novel algorithm perform in clinical practice? Would it significantly influence the designated targets of reduction in intraprocedural bleeding and procedure time? A complex randomized clinical trial would be required, in which unintended intraprocedural bleeding is one of the end points.

The study by Scheppach et al. demonstrates that intraprocedural bleeding during endoscopy can potentially be reduced in the future with the help of AI, but not completely vanish. Whether this will be an invaluable tool, as the authors suggest, requires confirmation in further studies, and these results will be awaited with great interest.

Publication History

Article published online:
14 April 2025

• References

• 1 Scheppach MW, Mendel R, Muzalyova A. et al. Use of artificial intelligence in submucosal vessel detection during third-space endoscopy. Endoscopy 2025; 57
• 2 Ebigbo A, Mendel R, Scheppach MW. et al. Vessel and tissue recognition during third-space endoscopy using a deep learning algorithm. Gut 2022; 71: 2388-2390
• 3 Saito I, Tsuji Y, Sakaguchi Y. et al. Complications related to gastric endoscopic submucosal dissection and their managements. Clin Endosc 2014; 47: 398-403
• 4 Messmann H, Bisschops R, Antonelli G. et al. Expected value of artificial intelligence in gastrointestinal endoscopy: European Society of Gastrointestinal Endoscopy (ESGE) Position Statement. Endoscopy 2022; 54: 1211-1231
• 5 Uraoka T, Igarashi M. Development and clinical usefulness of a unique red dichromatic imaging technology in gastrointestinal endoscopy: a narrative review. Ther Adv Gastroenterol 2022; 15: 1-11
• 6 Funasaka K, Horiguchi N, Yamada H. et al. Endoscopic submucosal dissection of early gastric cancer using a novel image-enhanced endoscopy: amber-red color imaging. Endoscopy 2024; 56: E640-E641