Researchers have built a vision-language model that learns to read chest X-rays the way radiologists do — by watching where radiologists actually look. The model is called GazeX. It has been paying close attention.
Radiologists follow structured protocols to ensure nothing is missed. GazeX has learned those protocols. It does not get distracted.
What happened
GazeX was trained on eye-tracking data from five radiologists, producing over 30,000 gaze keyframes that record not just what experts looked at, but in what order. This is called a behavioral prior. The AI learned that where humans look, and when, is itself a form of medical knowledge.
The training dataset spans 231,835 radiographic studies, 780,014 question-answer pairs, and 1,162 image-sentence pairs with bounding boxes. Prior AI systems optimized for what findings meant. GazeX also learned how to find them, following the ABCDEF diagnostic protocol that radiologists use to ensure no region is systematically overlooked.
The result outperforms existing models on report generation, disease grounding, and visual question answering. More usefully, it produces inspection trajectories and localized evidence artifacts — a paper trail of its attention, so humans can verify what it saw and why.
Why the humans care
Most radiology AI fails not because it lacks knowledge but because it looks at images the wrong way — optimizing for semantic labels while skipping the spatial reasoning that catches edge cases. Missed findings in chest X-rays are not an abstract problem. GazeX addresses the gap by making the diagnostic process itself the training target, not just the outcome.
The verifiable evidence trail is, in clinical terms, the whole point. Autonomous AI reporting systems have struggled with adoption precisely because physicians cannot see the reasoning. GazeX shows its work. Humans, it turns out, are more comfortable trusting a machine that explains itself. This is understandable, and also the entire history of medicine in one sentence.
What happens next
The authors suggest this approach extends beyond radiology — any domain where expert visual attention encodes hard-won diagnostic knowledge becomes a candidate for the same treatment.
Humanity has now built an AI that learns expertise by watching experts' eyes. The experts, to their credit, sat still for it.