Abstract
Contrast sensitivity (CS) is a fundamental measurement that reveals the capabilities and limitations of human vision. Understanding the stimulus dependence of CS provides links to the underlying neural mechanisms and supports applications ranging from display optimization to diagnosis of retinal disease. Our broad objective is to establish principles and computations that enable prediction of CS from models of the initial visual encoding. Here we consider how encoding by the midget retinal ganglion cells (mRGCs) shapes CS for achromatic (L+M+S) and chromatic (L-M) patterns, presented at multiple eccentricities. We modeled cosine-windowed gratings (2 deg field) presented at eccentricities 0, 2.5, and 7 deg, with spatial frequencies from 0.25 to 64 cpd. The visual scenes were converted into retinal images using eccentricity- and wavelength-dependent optical point spread functions, derived from published wavefront-aberration measurements. Then, we computed the cone excitations of a simulated retinal mosaic and passed these through a model of mRGC receptive fields (RFs). The spatial linear model RFs are constrained by published anatomical and physiological data, and are non-selective with respect to input from the L- and M-cones. Finally, we derived CS by training a classifier with stimulus-labeled mRGC responses and computing the accuracy of the trained classifier using a simulated psychophysical task. The derived CSs recapitulate core features of human performance: i) CS as a function of spatial frequency is bandpass for achromatic gratings but lowpass for chromatic gratings; ii) peak CS shifts to lower spatial frequencies with increasing eccentricity; iii) at low spatial-frequencies, chromatic CS declines more rapidly with eccentricity than achromatic CS. Our results confirm the value of investigating limits of visual performance imposed by known features of the initial visual encoding, and provide a framework for integrating diverse experimental data into a generalizable model of CS.