Abstract
Introduction: Modern medical, aerial and satellite imaging generate increasingly larger volumes of images. For medical imaging, the use of volumetric data has drastically increased, although there is no full understanding of its impact on human search. When compared to 2D imaging, 3D data increases the information by providing various slices of the signal's volume but also increases the search space and the spatiotemporal uncertainty of the signal's location. Here, we evaluated human vs. ideal observer performance in 2D and 3D search for different signals with different detectabilities in the human visual periphery. Methods: Seven observers searched for a larger 3D-Gaussian signal (0.50 deg for its central slice) or a smaller sharp-edged sphere (0.13 deg) embedded in 2D or 3D 1/f2.8 isotropic filtered noise. Human observers were cued to find (Yes/No task; 50% target prevalence) one of the two signals in the 3D volume or in a 2D slice (central slice). Results: Ideal observer performance (d') increased from 2D to 3D by a factor of ~12 for the sharp sphere and ~170 for the Gaussian, in d' units. In contrast, human search performance increased for the Gaussian signal (2D d'= 1.28 ± 0.14 3D d'= 2.8 ± 0.35; p < 0.01) but decreased for the smaller spherical signal (2D d'= 3.92 ± 0.311 3D d'= 2.3 ± 0.39; p < 0.01). Analysis of human gaze behavior shows an increase (7 times higher) in search errors (not foveated) for the small signal in 3D images suggesting that its lower detectability in the visual periphery mediates the decrease in 3D search performance. A foveated search model correctly predicts the lower detectability of the small signals in 3D images. We conclude that the foveated nature of human visual processing, not captured by the ideal observer, has important implications on the effectiveness of 3D search.
Meeting abstract presented at VSS 2018