Detecting Potential False-Normal Predictions in AI Analysis of Celiac-Related Duodenal Endoscopy Images

In the modern clinic, doctors often rely on a camera to peer inside the human body, capturing still images of the lining of the small intestine to look for signs of disease. One such condition, celiac disease, causes the tiny finger-like projections in the intestine to flatten and shrink, a change that can sometimes be spotted in these photographs. While a doctor can usually confirm a diagnosis by taking a tissue sample for laboratory testing, the visual clues in an image are valuable for guiding that decision. Recently, computer programs have been trained to look at these same images and decide whether they show a healthy intestine or one affected by celiac disease. The hope is that these programs can act as a second pair of eyes, catching subtle signs that a human might miss or flagging images that need a closer look. However, a computer is not perfect. Sometimes, it confidently declares an image to be normal when the label says it is actually diseased. The critical question for anyone using this technology is not just how often the computer gets it right, but how to find the specific images where it is wrong before they slip through the cracks.

A researcher set out to test a practical solution to this problem using a collection of 188 digital photographs of the duodenum, the first part of the small intestine. These images came from a public archive and were already labeled by experts as either showing a normal intestine or one with celiac disease. Before running any tests, the researcher carefully cleaned the data, removing five images that were exact digital copies of others and setting aside 24 images that were marked as "doubtful" because their condition was unclear. This left a final group of 159 distinct images to work with. The goal was to see if a computer program could be taught to spot its own mistakes. Specifically, the researcher wanted to know if adding simple checks for image quality and stability could help a human reviewer prioritize which "false normal" images to inspect first, compared to simply looking at how confident the computer was in its own answer.

To do this, the researcher trained a standard computer vision model, a type of software known as a ResNet-18, to distinguish between the two types of images. The training was done in a way that ensured the model had never seen the specific images it was later tested on, preventing it from simply memorizing the answers. Once the model was ready, it was asked to classify every image in the test set. The model correctly identified most of the images, but it made a specific kind of error: it labeled 13 images as "normal" when the original expert labels said they were "celiac." These 13 images became the target for the experiment. The researcher then created two different lists to see which method would catch more of these 13 hidden errors if a human could only review a small fraction of the total images.

The first method was the simplest possible approach: the researcher simply looked at how confident the computer was when it called an image "normal." If the computer was only slightly sure, that image was flagged for review. The second method tried to be smarter by adding extra signals. It looked at whether the computer's answer changed slightly if the image was made a bit brighter or darker, or if the file was saved with slightly lower quality. It also measured the image itself for technical flaws like blurriness, extreme brightness, or glare in the center of the photo. These three factors—weak confidence, unstable answers, and poor image quality—were combined into a single score to rank the images for review. The researcher then simulated a scenario where a human could only review a fixed number of images, such as 20% of the total, to see which method would find more of the 13 missed cases.

The results were clear and surprisingly straightforward. When the researcher reviewed the top 15 images flagged by the "confidence only" method, they found three of the 13 missed cases. When they reviewed the top 15 images flagged by the more complex "combined score" method, they also found exactly three of the 13 missed cases. The extra checks for image quality and stability did not help the system find more errors than just looking at the computer's confidence level. In fact, at a lower review rate where only a few images could be checked, the simple confidence method actually found more errors than the complex method. The researcher also tried a variation where the images were cropped to remove the edges and borders before being analyzed, hoping to see if the computer was relying on visual shortcuts like text or device overlays, but this change did not improve the performance of the complex scoring system either.

The study concludes that for this specific set of images, adding technical quality checks and stability tests did not improve the ability to catch the computer's mistakes. The extra complexity did not earn its keep. This does not mean that image quality or stability are unimportant in general, nor does it mean that the computer is perfect. It simply means that in this specific test, the computer's own sense of confidence was just as good at pointing out the risky images as a more complicated system that tried to measure the image's technical flaws. The researcher notes that the dataset was small, containing only 159 images, and the labels were based on image categories rather than confirmed patient biopsies, so the findings are limited to this specific archive. The work serves as a careful, transparent test showing that one should not assume a more complicated safety score is better than a simple confidence check without direct evidence. In the end, the most effective way to find the computer's errors in this case was to trust its own uncertainty, rather than trying to measure the picture itself.