Goals and objective
Eye tracking is commonly used for investigating expertise differences in radiology (Krupinski, 2010). Most expertise research focuses on detection of nodules (small, inconspicuous abnormalities) but also other types of diseases might be present in radiological images such as chest radiographs. Three types of images were compared, showing: (a) a focal disease (abnormality at a specific location), (b) a diffuse disease (involves all lobes of both lungs), and (c) no abnormalities (normal images). We investigated the eye movements of experts, intermediates, and novices on those images. It is known that eye movements, which reflect deployment of attention, can be influenced by bottom-up effects which result, for example, from the type of image, and top-down effects which result, for example, from the expertise of the viewer. Bottom-up and top-down effects interact and, together, they ‘guide’ attention. This interaction is commonly neglected in expertise research, but the current study investigates the interaction between the type of disease and expertise on eye movements.
Eleven sixth-year medical students (novices), ten residents (intermediates), and nine radiologists (experts) inspected 24 chest radiographs (8 focal, 8 diffuse, 8 normal images) and orally provided a diagnosis while their eye movements were recorded.
Eye tracking measures
An Eyelink 1000 remote eye tracker was used. A global/local ratio was computed by dividing the number of long saccades ( > 1.6 degrees of visual angle) by the number of short saccades (< 1.6 degrees of visual angle) (Zangemeister, Sherman, & Stark, 1995). A higher ratio reflects more dispersed looking, while a lower ratio reflects looking at specific regions. Multilevel analysis was used to analyze this global/local ratio and the average fixation duration.
Main outcomes and significance
On average fixation duration, multilevel analysis showed a significant effect of type of image, F(2, 23.3) = 9.2, p < .001; a significant interaction effect of type of image and expertise, F(4, 659.1) = 5.17 , p = .001, but no significant effect of expertise, F(2, 30.0) = 1.12, p = .34. On the global/local ratio, we found a main effect of type of image, F(2, 23.6) = 4.74, p = 0.02, and significant interaction of image type with expertise, F(4, 651.2) = 4.72, p = 0.001. No significant effect of expertise was found, F(2, 30.0) = 0.44, p = .65.
Type of images strongly influenced participants’ eye movements on radiological images. Regardless of expertise, in focal images participants looked relatively long at specific locations, while in diffuse images, relatively short, dispersed looking took place. For normal images, an expertise effect is visible: students looked more dispersed on diffuse compared to normal images, while residents and radiologists looked most dispersed on normal images. In conclusion, viewing behavior differences between experts and novices are relatively small on disease images but larger on normal images. However, students’ diagnostic accuracy on normal images is quite high, while for disease images, their accuracy is very low. Models of expertise effects on eye movements should take into account effects of image type.
Krupinski, E. A. (2010). Current perspectives in medical image perception. Attention Perception & Psychophysics, 72, 1205-1217.
Zangemeister, W. H., Sherman, K., & Stark, L. (1995). Evidence for a global scanpath strategy in viewing abstract compared with realistic images. Neuropsychologia, 33, 1009-1025.