GIM-ENDO: A Multimodal Endoscopic Image and Video Dataset for Gastric Intestinal Metaplasia Morphology and Pathology

Imagine the human body as a vast, bustling city, and the stomach as a busy marketplace where food is broken down. Sometimes, the walls of this marketplace get irritated and change their texture, turning from a smooth, pink pavement into something rough and patchy. This change is called "gastric intestinal metaplasia" (GIM). Think of it like the city's pavement starting to crack and sprout weeds; it's not the final disaster, but it's a warning sign that the ground is unstable and could eventually lead to something much worse, like a building collapse (cancer). For a long time, doctors have had to guess which patches of "weeds" are dangerous by looking at them through a camera, but it's hard to tell just by looking.

Enter Artificial Intelligence (AI), the super-smart detective that can learn to spot these dangerous weeds instantly. But here's the catch: AI is only as good as the training manual it reads. Until now, there hasn't been a really good, open-source manual that combines clear photos of the stomach, video tours of the area, and the "truth" from a lab test (biopsy) to prove what the AI is seeing. Without this perfect manual, the AI detective is stuck guessing. This paper introduces a new, free training manual called GIM-ENDO, designed to teach computers how to spot these early warning signs in the stomach so they can help doctors catch trouble before it becomes a crisis.

The New "Stomach Detective" Manual

The researchers behind this project, a team of doctors and data scientists from Iran, realized that while AI is getting better at spotting stomach issues, it's been held back by a lack of good data. They created GIM-ENDO, a special collection of images and videos that acts like a "gold standard" textbook for training computers.

Think of the stomach lining like a complex landscape. When it's healthy, it looks like a smooth, uniform field. When it has GIM, it starts to show specific "landmarks" or signs, like strange blue lines, cloudy bands, or white patches. The GIM-ENDO dataset is a treasure chest of 98 still photos and 39 video clips (each about 50 seconds long) taken from 24 patients. Most of these patients had the "weedy" GIM condition, while two were healthy controls.

What makes this dataset special is that it doesn't just show the pictures; it comes with a detailed "answer key." For every photo and video, the researchers have labeled exactly what the AI should be looking for. They identified six main "clues" that appear in the stomach when GIM is present:

• Light Blue Crest (LBC): A tiny blue line that looks like a ridge.
• Marginal Turbid Band (MTB): A cloudy, fuzzy edge around the cells.
• White Opaque Substance (WOS): A white, shiny patch.
• TV Pattern: A specific texture that looks a bit like a TV screen's static or a fusion of patterns.
• Atrophy: Where the stomach lining has thinned out.
• Map-like Erythema (MLE): Red patches that look like a map.

The team also recorded if the patients had H. pylori (a common stomach bacteria) and, crucially, they matched every image to a real lab test result. This means they know for a fact whether the "weeds" are the "complete" type (less dangerous) or the "incomplete" type (more dangerous), and they even noted the stage of the disease using systems called OLGA and OLGIM.

How They Built It

The team used a high-tech camera system (Olympus EVIS X1) to take these pictures. They didn't just snap a quick photo; they used special lighting modes (like Narrow-Band Imaging) that make the tiny details of the stomach lining pop out, almost like turning on a flashlight in a dark cave to see the cracks in the wall.

They carefully selected patients who were having stomach scopes for various reasons. If a patient had a suspicious spot, the doctors took a tiny sample (a biopsy) and sent it to a pathologist. The pathologist looked at the sample under a microscope to confirm the diagnosis. Only after the lab confirmed the "truth" did the team label the corresponding images and videos. This process ensures that the AI isn't just guessing; it's learning from facts that have been double-checked by experts.

What This Means for the Future

The authors are careful to say that this dataset is a starting point, not a finished solution. With only 24 patients, it's a small library, perfect for "pilot studies" or for researchers to test their ideas, but not big enough to train a super-advanced AI that can be used in every hospital tomorrow. They admit that because the group is small, the AI might need more practice with different types of people and stomachs before it's ready for the real world.

However, the potential is exciting. By giving researchers a clear, structured way to teach computers what these stomach signs look like, GIM-ENDO could help build AI tools that act as a "second pair of eyes" for doctors. Imagine a future where, during a routine check-up, an AI system instantly highlights the dangerous "weeds" on the screen, telling the doctor exactly where to take a biopsy. This could help catch stomach problems earlier, allowing for simpler treatments and potentially saving lives by stopping cancer before it even starts.

The dataset is now free for anyone to download and use for non-commercial research, with the hope that more scientists will join in to make the "detective" even smarter. As the team notes, this is just the first chapter in a story that could change how we fight stomach cancer.