CBAM-Enhanced MaxViT-Tiny for Gastrointestinal Endoscopy Image Classification: Robustness, Explainability, and External Validation

Inside the human body, the digestive tract is a long, winding tunnel that doctors must inspect to find signs of disease. To do this, they use a flexible tube with a camera, known as an endoscope, which captures thousands of images during a single procedure. These pictures show the lining of the stomach and intestines, revealing everything from normal folds to dangerous growths like polyps or inflamed tissue. While trained specialists can read these images, the sheer volume of data they generate makes it difficult to review every single frame without fatigue. This has led researchers to ask if artificial intelligence could help, acting as a tireless second pair of eyes to sort through the images and flag potential problems. The hope is that a computer program could learn to recognize the subtle visual patterns of disease, offering a reliable and fast way to assist in diagnosis.

However, there is a significant catch in trusting these computer systems. While they can be incredibly accurate when looking at clear, standard medical photos, they can be surprisingly fragile. If the image is altered even slightly by invisible noise or if the lighting changes, the computer might suddenly become confused and give a completely wrong answer. This vulnerability is a major concern for medical safety, as a system that fails when the conditions are not perfect cannot be trusted in a real hospital. Furthermore, even when a computer gets the right answer, it is often a mystery why it did so. It might be focusing on the actual disease, or it might be relying on a timestamp or a border on the screen that happens to be present in the pictures of sick patients. To be useful, an artificial intelligence needs to be not only accurate but also robust against these errors and transparent about how it makes its decisions.

A team of researchers set out to build and test a new type of artificial intelligence designed to classify these gastrointestinal images. They started with a powerful computer model known as MaxViT-Tiny, which is a hybrid system that combines two different ways of processing visual information: one that looks at small local details and another that understands the bigger picture. To help this model focus on the most important parts of an image, they added a specialized attention mechanism, a feature that acts like a spotlight, telling the computer to ignore irrelevant background noise and concentrate on the tissue itself. They trained this system on a large collection of nearly 4,000 endoscopic images, carefully removing duplicate pictures to ensure the data was clean. The images covered eight different categories, including normal landmarks like the opening of the stomach and the beginning of the large intestine, as well as various conditions like ulcers, inflammation, and polyps.

The researchers split these images into three groups: one for teaching the model, one for checking its progress, and a final group that the model had never seen before, to test its true skill. When they ran the test, the new system performed exceptionally well, correctly identifying the condition in more than 94 percent of the unseen images. It outperformed five other popular computer models that were tested in the same way, showing that this specific combination of architecture and attention was effective. The system was also very good at distinguishing between the different types of tissue, with a success rate that remained high across multiple test runs. This suggested that the model had genuinely learned to recognize the visual features of the diseases rather than just memorizing the training pictures.

But the researchers knew that high accuracy on a clean test was only half the story. They wanted to know if the system would hold up if someone tried to trick it. To find out, they subjected the model to a series of digital attacks, where they added tiny, invisible distortions to the images that are imperceptible to the human eye but can confuse a computer. The results were sobering. When faced with these tricky, manipulated images, the model's accuracy plummeted. In some cases, the system got the answer right less than four percent of the time. Even when they tried to retrain the model to be tougher against one specific type of trick, it remained vulnerable to other, more complex attacks. This revealed a critical weakness: the model was not robust enough for real-world use where it might encounter unexpected variations or deliberate interference.

The team also looked inside the "mind" of the computer to see what it was actually looking at when it made a decision. They used visualization tools to draw heat maps over the images, highlighting the areas the model considered most important. In many cases, the system correctly focused on the diseased tissue or the anatomical landmarks. However, in other instances, the model was distracted by non-medical artifacts, such as the text overlays, the borders of the screen, or the time stamps printed on the images by the camera. This meant the computer was sometimes using shortcuts, relying on clues that would not be present in a different hospital or with a different camera. When the researchers tested the model on a completely separate set of images from a different source, its performance dropped noticeably, confirming that it had not learned to generalize as well as hoped.

Ultimately, the study concluded that while this new artificial intelligence model is a strong performer on standard, clean images, it is not yet ready for the clinical environment. The high accuracy seen in the lab did not translate into safety against digital tricks, and the model's reliance on image artifacts suggests it could fail in different settings. The researchers emphasize that achieving a high score on a test is not the same as building a reliable medical tool. Before such a system can be used to help doctors diagnose patients, it must be made much more robust against errors and trained to ignore irrelevant details. The work serves as a necessary reality check, showing that the path to trustworthy medical artificial intelligence requires more than just high accuracy; it demands resilience, transparency, and validation across diverse real-world conditions.