Anqi Zhang¹ (), Miguel P. Eckstein¹; ¹Department of Psychological and Brain Sciences, UC Santa Barbara

Introduction: Radiologists regularly search for disease in slices acquired by 3D imaging techniques. Their navigation strategy varies (Drew et al., 2013; Aizenman et al., 2017; Ba et al., 2020). Some scroll through slices while maintaining fixation (drilling), whereas others deploy multiple eye movements within one slice before advancing (scanning). However, it remains unknown when and why either strategy is more effective than the other, and what the performance-maximizing strategies might be. Methods: We derived an image-computable Foveated Ideal Searcher in 3D (3D-FIS) that accounts for foveated visual processing and noise correlations across fixations and slices. We compared the 3D-FIS to model searchers with systematic drilling and scanning for a larger target visible in the periphery (mimicking a mass in breast images) and a smaller target difficult to detect in the periphery (mimicking a microcalcification). We assessed localization accuracy of human observers (N=12) for each target using drilling and scanning strategies with gaze monitoring and auto-scrolling display. On each search trial, observers gave an early localization response after the 21st fixation and a late response after the 80th. Results: The 3D-FIS shows emergent drilling predominance for the larger target, and alternates scanning and drilling for the smaller target. The drilling model approximates 3D-FIS accuracy for the larger target. The scanning model attains closer accuracy to the 3D-FIS than the drilling model for the smaller target. These results predict our psychophysical findings. With short search time, human drilling outperformed scanning for the larger target; with long search time, human scanning outperformed drilling for the smaller target (both p < .0001). Furthermore, the foveated models with recorded eye movements yield the same pattern of performance differences. Conclusion: We establish a new theoretical foundation for understanding optimal visual search strategies in 3D volumes, explaining radiologists’ search behavior, and predicting the effectiveness of heuristic strategies.

Acknowledgements: Supported by NIH grant R01EB026427.