Development and Validation of a Feature-Prompted Deep Learning System for Diagnosing Chronic Atrophic Gastritis Under White-Light Endoscopy

Every year, millions of people around the world face the threat of stomach cancer, a disease that remains one of the leading causes of cancer-related death globally. Before this cancer develops, the stomach lining often undergoes a slow, silent transformation known as chronic atrophic gastritis. In this condition, the healthy tissue that lines the stomach gradually thins and loses its function, a change that significantly raises the risk of cancer appearing later. Detecting this early warning sign is crucial, yet it remains a difficult task for doctors. The primary tool for looking inside the stomach is a flexible camera on a thin tube, which captures images of the organ's interior. However, identifying the subtle signs of tissue thinning in these images is notoriously tricky. The changes can be faint, and even experienced specialists often disagree on what they are seeing or miss the signs entirely. Because of this uncertainty, doctors sometimes rely on taking random tissue samples from various parts of the stomach to be sure, a process that is time-consuming, costly, and uncomfortable for patients.

A team of researchers at the Hangzhou Xiaoshan Hospital of Traditional Chinese Medicine has developed a new artificial intelligence system designed to help doctors spot these early warning signs more reliably. Their goal was to create a digital assistant that could look at the same live video images a doctor sees during a standard examination and identify the specific patterns of tissue thinning with high accuracy. Unlike previous attempts that tried to analyze a single image in one go, this team built a multi-step system that mimics the careful way a human expert examines a patient. The system first checks if the image is clear enough to be useful, discarding blurry or poorly lit frames. It then identifies exactly which part of the stomach the camera is looking at, distinguishing between the entrance, the main body, and the exit. Finally, it focuses on the specific areas where tissue thinning might be hiding, using a technique that highlights the suspicious regions before making a diagnosis.

The researchers tested this system using thousands of images taken from nearly one thousand patients who had undergone gastroscopy between January 2024 and September 2025. To ensure the system was learning the right things, they trained it on a vast collection of images that had been carefully labeled by experienced doctors. The most innovative part of their approach was how the system learned to find the disease. Instead of asking the computer to guess where the problem was in a whole picture, they first taught a separate part of the system to draw a digital outline around the areas that looked like they might be thinning. The main diagnosis tool then used these outlines as a guide, focusing its attention only on the relevant tissue. This method, which the researchers call feature-prompted learning, allowed the system to ignore the healthy parts of the stomach that might distract it, much like a person might focus on a specific detail in a crowded room while ignoring the background.

When the team put their best model to the test, the results were striking. The system correctly identified the presence of tissue thinning in 92.4% of the cases where it was actually there, and it correctly ruled it out in 89.2% of the cases where it was absent. This level of accuracy was not just a statistical success; it was a practical one. When the researchers compared the computer's performance against a group of human doctors, the system performed on par with the most senior specialists, who had more than five years of independent experience. The system's accuracy of 90.9% was statistically indistinguishable from the 90.3% achieved by these experts. However, the system significantly outperformed doctors with intermediate experience, who had been practicing independently for three to five years. These less experienced doctors achieved an accuracy of 85.6%, a gap that suggests the artificial intelligence could serve as a powerful equalizer, helping less experienced practitioners reach the diagnostic level of the most seasoned experts.

The study also revealed that the multi-step design of the system was essential to its success. When the researchers tried a simpler version that looked at the whole image without first filtering for quality or location, the accuracy dropped to 87.0%. This difference proved that the system's ability to clean up the data and focus on the right anatomical areas made a real, measurable difference in its final judgment. The feature-prompted strategy, where the system was guided by the digital outlines of the suspicious areas, further improved the results, showing that directing the computer's attention to the most relevant parts of the image helped it avoid mistakes. While the system showed great promise, the researchers were careful to note that this was a single-center study, meaning it was tested on patients from one hospital using specific equipment. They acknowledged that the system still needs to be tested in a wider variety of settings and with different types of cameras before it can be used routinely in clinics everywhere.

Despite these limitations, the work represents a significant step forward in the effort to standardize the detection of precancerous stomach conditions. By combining a careful, step-by-step analysis with a method that guides the computer to focus on the most important details, the researchers have created a tool that matches the performance of top-tier human experts. This system does not replace the doctor but offers a second pair of eyes that never gets tired and never misses a subtle clue due to fatigue or distraction. As the technology moves toward larger, multi-center trials, it holds the potential to make the early detection of stomach cancer more consistent and accessible, ensuring that the warning signs of chronic atrophic gastritis are seen clearly, no matter who is holding the camera.