Studying individual differences in eye-movements and application to tumor-detection, baggage-screening and other real-world vi

Studying individual differences in eye-movements and application to real-world search tasks.

Deborah J . Aks<Working draft -- 8/27/06>

The broad goal of these visual search studies is to learn more about the mechanism driving a volitional form of visual search such as when a person is actively searching a complicated scene for ambiguous or ill-defined information. Two such tasks include screening x-rays for tumors in the case of radiological diagnosis, and potential weapons in the timely case of airport security. Search usually appears haphazard in these and many other complex search tasks[1], but recent theoretical work shows subtle patterns revealing important properties of the underlying process[3]. A key objective of these studies is to evaluate human eye movements in these challenging search tasks (those that demand our attention), compare these data w/ model output & characterize which search strategies are effective.

X-ray screening. Given recent and ongoing terror alerts-- some real, some false -- most would agree we face major challenges in effectively detecting security threats. The little research there is on baggage screening point to a similar conclusion on the significant limits we now face in accurate and efficient detection of weapons[4]. While some solutions are being met by advances in automated devices (for detecting well-defined targets)[5], or expanding our arsenal of bomb-sniffing canines (for chemical based weapons), humans --despite their many other limitations-- still excel in their pattern recognition skills especially in the visual realm when targets are camouflaged among a cluttered array of items sharing similar properties[6].

What have we learned from human research?

McCarley et al (2004) have focused on whether screening can be improved throught practice [7]. Their results, along with more basic research studies, show the tremendous difficulty improving search performance when the target is uncertain & when the context changes[8]. Standardizing how travelers pack their suitcases would offer one way to improve weapons detection; but, at this point in time, most would agree the cost would be too extreme if not frightening in its implications, Instead of focusing on improving search through practice or luggage standardization, I have in mind looking at individual difference in search performance. The field of radiology has plenty of formal evidence documenting the wide range of search abilities among radiologists[9]. There are also the many anecdotal reports among radiologist and also airport screeners. Although this tremendous variability in search abilities s widely recognized, we still do not know what aspects of search makes certain people effective at detecting targeted information. Interestingly, intelligence seems unrelated to search skills [10].I believe we might make some progress by analyzing how individual search dynamic differs across people. Evaluating varous parameters of eye-movements might yield some progress. Some examples include analysis of:

Probability distributions: Does the degree of skew correlate w/ search performance?
Power spectra: Is more effective search best characterized by 1/f or some other spectral pattern? i.e., what color of noise?
Hurst exponent: (Does H correlate with search performance? Does better search have a persistent, anti-persistent or non-persistence across eye-movement sequence?
Additional tests to assess whether there is structure in these seemingly random search patterns can be found here.

Do changes in search across conditions reveal adaptive properties of effective search?
A crtical aspect of this project will evaluate the impact of systematic manipulation of various conditions (e.g., the degree of background structure) on search. If any correlations emerge do they reveal important properties of search -- such as the system's ability to adapt? Further evaluation of whether we see similar effects in search modeling will help clarify both underlying mechanism and whether such (adaptive) properties are critical to effective search.

The models
The ubiquity of 1/f & other power laws in biology [11] & now human eye-movements hints at an essential feature of these systems. Progress in improving automated and perhaps human search is likely to occur if we focus on models that produce key characteristics of this behavior: For example the Cellular Automata (CA) approach (e.g., Per Bak’s SOC model, predator/prey models such as discrete Volterra, Levy flight models and a variety of diffusion and other models). Considering only biologically plausible models will help narrow the range of possibilities (which means veering away from ARMA models that are so popular in cognitive and other sciences).

Bridging models of visual system with those that recognize the power law nature of behavior
Itti & Koch 2001 review key aspects known properties the human visual system The 2 stage (dorsal/ventral) model & the property of re-entry (i.e., iterative feedback) is widely accepted as essential features of the human visual system. Hamker (2003) incorporates into these ventral/dorsal streams the important property of ‘reentry’ and delineates the key regions involved in vision tasks (i.e., FEF, IT, V4). Both of these models though, like the bulk of vision theory-- do not address the power law nature of visual processes.Howevery if we look to other disciplines such as in biological and ecological sciences which focus on animal behavior we see wide-spread use of models that do consider the pervaisve power-law nature of animal behavior. Benichou et al (2005) for example, describe a 2 phase diffusion model (using Chapman-Kolmogorov differential equation). They describe a high diffusion model for fast search ,and slow diffusion model for scrutinizing local regions of search. While the modeling here describes search behavior in foraging animals it is clearly applicable to modeling human eye-movements. The relevance becomes more obvious as we see the strong parallels across these coarser animal behaviors with finer eye-movement behaviors used in visual search: the intermittency, the power laws & the efficiency of the animal behavior is very similar to human search. Recognizing the parallels across what appear to be very different domains of study, is likely to offer useful solutions to modeling effective search. These ideas will be elaborated in future drafts of this work.

Footnotes & References

[1].ref[2] ref[3].ref[4] ref[5].ref[6] ref

[7] McCarley, J.S., Kramer, A.F., Wickens, C.D., Vidoni, E.D. & Boot, W.R. (2004). Visual skills in airport-security screening. Psychological Science, 15,5, 302-306

.[8]Wolfe

[9].ref[10] ref[11].ref[12] ref

Benichou, O., Coppey, M., Moreau,M., Suet, P-H. & Voituriez, R. (2005). Optimal search strategies for hidden targets. Physical Review Letters, 94, 198101-04.

Hamker, F. (2003). The reentry hypothesis: linking eye movements to visual perception. Journal of Vision, 3, 808-816.

[x] .. those best identified by their visual features, such as sharp edges,