TopoGaze: Topological Analysis of Radiologist Gaze Patterns for Thoracic Disease Classification

Radiological image interpretation is cognitively demanding, and radiologist eye gaze offers a unique window into this process by capturing how attention shifts across anatomical and pathological regions. While prior studies link gaze patterns to diagnostic accuracy and expertise, analyses often reduce gaze to coarse metrics, discarding richer spatial structure. We introduce a novel framework applying topological data analysis (TDA) to model gaze as point clouds and graphs, from which we extract graph-theoretic descriptors and persistent homology features. Using the REFLACX dataset of chest radiographs with associated gaze and confidence annotations, we compute multi-scale topological signatures for each radiologist–image pair and train classifiers to predict thoracic diseases. Our model achieves strong performance (F1 = 0.748, recall = 0.856), with informative features including average node degree, Laplacian energy, spectral gap, and H₁ persistence. We further show that gaze topology varies systematically with diagnostic confidence, with significant differences in spectral and Betti features between high- and low-confidence cases. These results demonstrate that gaze topologies encode latent cognitive strategies and provide complementary predictive value for disease classification. To our knowledge, this is the first application of TDA to radiologist gaze, establishing behavioral topology as an interpretable biomarker for diagnostic reasoning.