A World Model of Radiologist Reading for Medical Image Representation Learning
Researchers have developed GazeWorld, a novel world model for medical imaging that learns from radiologist eye-tracking data. This model treats the image as a world and the radiologist's gaze sequence as a trajectory, autoregressively predicting representations of fixated image patches. When used as a pretraining paradigm, GazeWorld features achieve state-of-the-art diagnostic accuracy on multiple benchmarks, outperforming existing methods even without explicit gaze prediction training. AI
IMPACT This research demonstrates a new pretraining paradigm for medical imaging AI by modeling expert reading behavior, potentially improving diagnostic accuracy and zero-shot performance.