Speeding up visual field tests by incorporating spatial models

Abstract

Purpose: Many perimetric test algorithms exploit spatial relationships in the visual field (VF) for the initial location estimates of sensitivities or in a post-processing phase. We define a new algorithm (LiSZE: Likelihood Scaling ZEST) that uses spatial information on every presentation to alter VF estimates. We hypothesised that LiSZE would reduce test times without detriment to output precision and accuracy.

Methods: LiSZE is a maximum likelihood Bayesian procedure, which maintains a separate probability mass function (PMF) for each VF location. For each presentation, the location whose PMF has the largest standard deviation is tested, with the stimulus level set at the mean of the PMF. The PMF is updated with a likelihood function whose shape is dependent on the observer’s response. Any location in the VF that is related to that tested - according to a spatial model - is also updated with a modified likelihood function (scaled vertically and translated via an eccentricity correction factor). Spatial models were created based on: empirical data (Gardiner et al, Vis Res 2004), computational models (Denniss et al, IOVS 2012), nearest neighbour and random relationships.<br /> <br /> LiSZE was tested using computer simulations on 163 glaucomatous and 233 normal VFs (HFA 24-2 FT). ZEST was simulated as a baseline. Output measures included: number of presentations and visual sensitivity estimates. Errors were calculated by subtracting the estimated VF from the input VF.

Results: Median error for each VF was used as a global index of accuracy. LiSZE had similar accuracy to ZEST (median normal(dB)/glaucoma(dB): ZEST 0.5/0, LiSZE 0/-0.5-0), but higher variability (5^th-95^th percentile normal(dB)/glaucoma(dB): ZEST 1/1, LiSZE 2/1-1.5). We split results by input sensitivity (IS) to reveal differences between the LiSZE spatial models. The random model performed worst (absolute error and IQR worse than ZEST by up to 6dB for IS <28dB). The empirical model was best, with similar accuracy and better precision than ZEST (most IS had 1-2dB greater IQR (variability) for ZEST than empirical model LiSZE). Inspection of VF maps showed that LiSZE was able to detect localised VF loss. LiSZE was faster than ZEST: median number of presentations reduced by 10-25% for glaucoma and 25-46% for normals.

Conclusions: LiSZE has the potential to reduce VF test times. Simulations suggest that the empirical model produces a similar error profile to ZEST, while reducing test time.