GazeXPErT: An Expert Eye-tracking Dataset for Interpretable and Explainable AI in Oncologic FDG-PET/CT Scans

Eye-tracking data can improve AI transparency in medical imaging: The study found that AI models trained on expert eye-tracking data achieved better performance (Dice score of 0.6819) compared to models without gaze information (0.6008), suggesting that capturing where doctors look provides valuable signals for building more interpretable AI systems.

Expert gaze patterns reveal clinical decision-making: By recording 3,948 minutes of eye-tracking data from radiologists reviewing cancer scans, the researchers documented 9,030 gaze-to-lesion trajectories that capture how experts visually reason through diagnostic decisions—information that could help AI systems explain their recommendations in ways clinicians understand.

Natural human interactions could replace traditional computer interfaces in radiology: Rather than relying on keyboards and mice, the research proposes using eye-tracking as a more intuitive way for radiologists to interact with AI diagnostic tools, potentially streamlining clinical workflows and reducing cognitive burden.

AI can predict expert intent from gaze patterns alone: The baseline experiments demonstrated that an AI model could predict what a radiologist intended to do next with 67.53% accuracy based solely on their eye movements, suggesting gaze contains meaningful clinical information that could enable more responsive, adaptive AI assistants.

This dataset addresses a critical barrier to AI adoption in oncology: By creating a resource that links expert visual reasoning to AI decision-making, the work tackles the explainability and trust issues that have hindered clinical translation of standalone AI lesion detection models.