New AI Reads Head and Neck Cancer Scans Better Than ChatGPT

Doctors use PET/CT scans to find head and neck cancer, but the area is crowded with small, complex structures. Even experts say it's tough and time-consuming. To help, researchers in South Korea built a specialized AI model trained on over 40,000 scan-related question-and-answer examples, created by radiologists from real cases. This AI learned not only to spot whether a tumor is present, but also to locate which lymph nodes in the neck are affected.

When put to the test, the specialized model crushed general-purpose AI systems like ChatGPT. On questions about lymph node spread, the general models scored close to zero, while the new model scored high on accuracy and relevance. It also correctly identified the presence of the primary tumor about 83% of the time on scans it had trained on, and 69% of the time on completely new scans from four other hospitals. That gap is expected, since new machines and patients can differ, but it shows the approach works well enough to build on.

What does that mean for you? If you or a loved one ever needs a head and neck cancer workup, an AI like this could act as a tireless second set of eyes, pointing out suspicious spots that might be easy to miss at 5 p.m. after a long shift. It could also help train future doctors by explaining scan findings conversationally, like an intelligent tutor. The catch: the 69% accuracy on external data means it isn't ready to make decisions alone — and the researchers note that privacy and safety concerns around medical AI are still very real. For now, think of it as a promising assistant that must stay under a specialist's watch.