Purchase this article with an account.
Luis Andres Lesmes, Zhong-Lin Lu, Jongsoo Baek, Thomas Albright; Efficient adaptive measurement and classification of contrast sensitivity functions. Journal of Vision 2008;8(6):939. doi: 10.1167/8.6.939.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Purpose. The contrast sensitivity function (CSF), describing observer's grating sensitivity as a function of spatial frequency1, is a canonical measure of spatial vision. Clinically, a number of visual neuro-pathologies exhibit characteristic CSF deficits2. The cards/charts currently used for clinical testing, though easy to use, limit the sampling range and grain of grating contrast and spatial frequency, and therefore limit test precision.We sought to develop adaptive testing methods that (1) estimate CSFs with the precision of psychophysical testing and short testing time of cards/charts, and (2) classify CSFs into pathological categories based on candidate CSFs with very short testing time. Method. Describing the CSF with a simple functional form5, the quick CSF (qCSF) method searches one-step-ahead for the stimulus contrast and spatial frequency minimizing the expected entropy3–5 of the posterior jointly defined over four parameters: (1) peak sensitivity, (2) peak spatial frequency, (3) bandwidth at half-peak sensitivity, and (4) low spatial frequency truncation. In an orientation discrimination task, observers ran the qCSF concurrently with an adaptive method3 estimating thresholds independently at 6 spatial frequencies. Each of four sessions provided two qCSF estimates and one conventional CSF estimate. For adaptive classification4, candidates included one normal CSF, one CSF with general-deficit, and three CSFs with low, middle and high frequency-specific deficits. Stimulus placement minimized the expected entropy for the probability of class membership. Results. In agreement with simulations, psychophysical results validated that CSFs obtained with 30, 50, and 100 qCSF trials agree with conventional CSF estimates: (mean r=.944, .960, and .976). Classification simulations showed that detecting abnormal CSFs typically took less than 10 trials and classifying specific abnormal CSFs took 20–30 trials. These adaptive Bayesian methods, generating efficient and precise CSF estimates, have clear implications for laboratory and clinical applications. 1.Campbell & Robson(1968) 2.Regan(1993) 3.Kontsevich, and Tyler (1999) 4.Cobo-Lewis(1996). 5.Lesmes et-al(2006) 6.Watson and Ahumada(2005)
This PDF is available to Subscribers Only