Spatial artificial intelligence in gastric endoscopy for premalignant lesions and early gastric cancer: a systematic review
This systematic review of 39 study families highlights that while spatial AI in gastric endoscopy is advancing toward clinically relevant mapping of premalignant lesions and early gastric cancer, its readiness to support procedural decisions is currently limited by a lack of robust reference standards, prospective validation, and live evaluation.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Inside the human stomach, the lining is a delicate landscape where cancer can begin as a subtle, almost invisible change. For doctors using a flexible camera to look inside, the challenge has long been not just spotting a suspicious spot, but knowing exactly where it starts and where it ends. This boundary is critical. If a lesion is too small or too large, or if its edges are unclear, the treatment plan changes completely. Sometimes, a precise map allows a doctor to remove the cancer entirely with a camera-based procedure; other times, the uncertainty means a patient needs more invasive surgery. For years, computer programs have been trained to act as a second pair of eyes, helping to flag these dangerous areas. But a new question has emerged: can these programs do more than just say "something is wrong here"? Can they draw the exact outline of the problem with enough accuracy to guide a surgeon's hand?
A team of researchers set out to answer this by reviewing a large collection of studies that tested artificial intelligence systems designed specifically to map these stomach lesions. They looked at dozens of projects where computers were asked to trace the borders of early stomach cancer or pre-cancerous changes, rather than simply detecting their presence. The researchers found that while these systems have become quite good at drawing lines on stored images, the proof that they can do this reliably during a real, live medical procedure is still very thin. The technology has moved from simple detection to complex mapping, but it has not yet fully crossed the bridge from a promising tool in a lab to a trusted guide in a hospital.
The review examined thirty-nine groups of studies, mostly conducted in East and Southeast Asia, where stomach cancer is a significant health concern. These studies focused on two main targets: early stomach cancer and pre-cancerous conditions like intestinal metaplasia, which are known to increase cancer risk. The researchers looked at how these artificial intelligence systems performed when asked to segment, or separate, the abnormal tissue from the healthy tissue, and how well they could identify the specific edges of a lesion. In many cases, the systems were tested on thousands of still images taken from previous procedures. The results showed that on these stored pictures, the computers could often draw outlines that matched the drawings of human experts with a high degree of overlap. Some systems even managed to distinguish between different types of tissue or identify the exact line where a cancer ends and healthy tissue begins.
However, the researchers found a significant gap between these impressive numbers on paper and the reality of clinical practice. Most of the studies relied on images that had been saved and reviewed after the fact, rather than testing the systems while a doctor was actually performing an endoscopy. When the researchers looked for evidence of the technology being used in real-time, they found very little. Only two studies reported evaluating the system while the camera was moving inside a patient, and only one study showed that the computer's map actually changed what the doctor did during the procedure, such as prompting an extra biopsy. Furthermore, the way these systems were tested often lacked the rigor needed for medical safety. Many studies did not separate the data from different patients clearly, meaning the computer might have been tested on images from the same person it was trained on, which inflates the apparent success. Only a handful of studies tested the systems on completely new groups of patients from different hospitals, a crucial step to prove the technology works for everyone, not just the specific group it was built on.
The definition of "correct" also varied widely across the studies, making it hard to compare them directly. In some cases, the computer was judged against a map drawn by a doctor on a screen, while in others, it was compared to the actual tissue removed and examined under a microscope. The researchers noted that a doctor's drawing on a screen is not the same as the true biological boundary found in a lab. A system might match a doctor's drawing perfectly but still miss the true edge of the disease, which is the only edge that matters for a successful surgery. Because the standards for what counted as a "good" map were so different, the researchers could not combine the results into a single average score. Instead, they observed a landscape of varied performance where some systems worked well for specific tasks on specific images, but none had yet proven they could consistently define the limits of a lesion with the reliability required for life-or-death decisions.
The review concludes that while spatial artificial intelligence has made remarkable strides in understanding the geography of stomach disease, it is not yet ready to be the primary guide for treatment planning. The technology has successfully moved beyond simple detection to the complex task of mapping, but the evidence supporting its use in live procedures is still limited. The researchers emphasize that for these tools to become standard, future studies must test them on live patients, use stricter methods to ensure the computer has not simply memorized the answers, and prove that the maps they draw actually lead to better patient outcomes. Until then, these systems remain powerful assistants that can highlight potential trouble spots, but the final decision on where to cut or where to biopsy must still rest with the human doctor. The journey from a computer that can draw a line on a photo to a system that can safely guide a surgeon through a complex procedure is still underway.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.