Abstract
Purpose :
The qCSF method applies Bayesian active learning to provide an accurate, precise and efficient assessment of spatial vision (Lesmes et al. 2010). To date, qCSF testing has not been informed by regularities in CSF shape observed when individuals are tested across low, medium, and high luminance conditions. To improve CSF analysis, and leverage information provided by cross-test regularities, we developed a hierarchical Bayesian model (HBM), which infers joint posterior distributions of CSF parameters and hyperparameters from qCSF data obtained from 112 subjects tested in three luminance conditions (Hou et al. 2016).
Methods :
The CSF was modeled with a log-parabola with peak gain (PG), peak spatial frequency (PF), and bandwidth at half height (BH). The two-level HBM consisted of multiple 3-dimensional Gaussian distributions of CSF parameters at the population and individual test levels. The 3×3 covariance distributions at two levels quantified cross- and within-test regularities. The means of the parameter distributions at the individual test level were sampled from the hyperparameter distribution at the population level, while all individual tests shared the same 3×3 within-test covariance. We compared the average half-width of the 68.2% credible intervals (HWCIs) of the CSF parameters and area under log CSF (AULCSF) estimates with the qCSF and HBM.
Results :
The HBM recovered significant correlations among CSF parameters at the population (Fig. 1; r(PG&PF)=0.441, r(PG&BH)=0.580, r(PF&BH)=-0.109) and individual (r(PF&BH)=-0.719) test levels. The average HWCI (in log10 units) of the estimated CSF parameters and AULCSF decreased with the number of trials in both the qCSF and HBM analyses (Table 1). Analysis of AULCSF estimates obtained with 50 trials provided HWCI values of 0.040 for qCSF and 0.035 for HBM. Relative to estimates of CSF parameters and AULCSF obtained with the qCSF, the HBM reduced the HWCI by 60-74% and 32% with 15 trials, and 30-55% and 13% for 50 trials. The average absolute difference between qCSF and HBM estimates was not statistically significant.
Conclusions :
Incorporating both cross- and within-test regularities, the HBM can further improve the precision of CSF and AULCSF estimates, especially when the number of tested trials is relatively small.
This is a 2021 ARVO Annual Meeting abstract.