Stomach cancer: an AI spots lesions in endoscopy images in under a minute

Why is early stomach cancer hard to see?

An upper endoscopy is an exam where a doctor passes a thin tube with a camera through the mouth to look inside the stomach. It is the main way to find stomach cancer early, while it is still shallow and easier to treat.

The problem is that early cancer often looks a lot like gastritis, a common irritation of the stomach lining. Spotting small lesions takes years of experience, and each exam produces many images to review. The authors wanted to know if an AI could sort through these images automatically and ease the workload of endoscopists, the doctors who perform the exam.

What did the researchers do?

The team in Japan built a convolutional neural network, a type of AI designed to read images. They trained it on 13,584 endoscopy images of stomach cancer so it could learn to draw a box around suspicious areas.

To test it, they used a separate set of images the AI had never seen: 2,296 stomach images from 69 patients in a row, who together had 77 cancer lesions. The AI's answers were then checked against the known diagnosis for each lesion.

What were the results?

The AI analyzed all 2,296 images in 47 seconds. It correctly found 71 of the 77 cancers, a sensitivity of 92.2%. Sensitivity means the share of real cancers a test manages to catch. It found 70 of the 71 lesions measuring 6 millimeters or more (98.6%), and every invasive cancer, meaning one that had started to grow deeper into the stomach wall.

The 6 missed cancers were small, very shallow lesions. The authors describe them as hard to tell apart from gastritis, even for experienced endoscopists.

The weak spot was false alarms. The AI flagged 161 areas that were not cancer. As a result, only 30.6% of its alerts were true cancers. Nearly half of these false alarms were gastritis with changes in color or an uneven surface.

What could this mean for patients?

The authors see the tool as a helper for doctors: a very fast second pair of eyes that can review many images and point out areas worth a closer look. It is not meant to replace the endoscopist, who still makes the call and takes tissue samples (biopsies).

For patients, this could one day lower the chance that a small cancer is missed. But with this many false alarms, the tool would need to improve so it does not lead to too many unnecessary biopsies.

What are the limits of this study?

The AI was tested on stored, selected images, not live during an endoscopy. So we do not yet know how it performs in real practice, with blurry frames, glare, or movement.

The test included only 69 patients from one image set. False alarms remained high. Finally, the study does not show that any patient received better care because of the tool. That would require prospective studies carried out during real exams.