POEM uses endoscopic tools in the submucosal space at the esophogastric junction to alleviate the effects of achalasia.
Drs. Daniel A. Hashimoto, Gregory Ginsberg, Galen Leung, and Eric Eaton, representing GI Surgery, Gastroenterology, General Robotics, and Computer and Information Sciences at Penn Medicine, are developing a study that will incorporate artificial intelligence (AI) into per oral endoscopic myotomy (POEM) and endoscopic submucosal dissection (ESD) to augment human visual perception during these surgeries.
Limited perceptual field (e.g., lack of depth perception, narrowed vision, reliance on video vs. direct visualization) presents significant challenges in performing POEM and ESD. These "third space" surgeries are increasingly applied to the management of achalasia and early-stage GI cancers because they offer patients a less invasive alternative for conditions that previously required open or laparoscopic procedures. Third space (or intramural) endoscopy involves tunneling in the submucosal space to access the deeper layers of the gastrointestinal tract while maintaining the integrity of the overlying mucosa. [1]
Other challenges are common to third space endoscopy. While POEM and ESD employ similar techniques to achieve different clinical goals, for example, they share similar risks — including inadvertent injury causing perforation, leak, and subsequent mediastinal or intra-abdominal sepsis. In addition, the learning curve for POEM and ESD (between 100-250 cases) limits their adoption to specially trained interventional endoscopists. When compounded with the relatively higher adverse event rate and a dearth of training opportunities, these factors may explain the lower incidence of POEM and ESD in the United States by comparison to countries such as Japan.
An AI Answer in Sight
Recent advances in computer vision and AI may reduce the risks and potential technical errors identified with third space endoscopy by enabling automated detection of safe and unsafe areas of submucosal dissection and inadvertent injuries. Furthermore, computer vision guidance may alleviate the steep learning curve of these surgeries, making them more accessible to endoscopists and, thus, to more patients.
With these aims in mind, the study in development at Penn Medicine will first collect up to 500 POEM and ESD videos annotated for anatomy, safe and unsafe zones of dissection, tool-tissue interactions, and inadvertent injuries such as bleeding, thermal injury, or perforation. To determine a consensus on safe and unsafe areas of dissection, the videos will then be compared and analyzed via deep learning to identify target zones, anatomy, and tool interactions.
The need for vast amounts of data, or big data, is both a boon and barrier in AI. While big data is the fuel for AI, its massive, complicated, and high velocity datasets can be unwieldy, difficult to store and manage, and insecure. Thus, a further goal for the study will be the application of sophisticated algorithms to reduce the need for large amounts of data.
The study will take at least two years, and will be the precedent for future investigations evaluating whether inadvertent injuries detected by AI correlate with worsened outcomes such as post procedural leaks, infection, delay in oral intake/tolerance, increased length of stay, cancer recurrence (for ESD) or increased number of radiologic studies.
About the Study Faculty
Dr. Hashimoto is the Director of the Penn Computer Assisted Surgery and Outcomes (PCASO) Laboratory; Dr. Ginsberg is the Director, Penn Endoscopic Services; Dr. Leung is an Assistant Professor of Clinical Medicine at Penn Gastroenterology; Dr. Eaton is a faculty member in the Department of Computer and Information Science at the University of Pennsylvania, and a member of the General Robotics Automation, Sensing, Perception Lab.
Referral information for Penn GI Surgery and Penn Gastroenterology and Hepatology is available by calling 877-937-7366.
Reference
Shanbhag AB, Thota PN, Sanaka MR. Recent advances in third space or intramural endoscopy. World J Gastrointest Endosc 2020; 12(12): 521-531 [PMID: 33362905 DOI: 10.4253/wjge.v12.i12.521]