You are on page 1of 1

Using Structural Image Data to Seed a Perimetric Test Strategy for Glaucoma

Jonathan Denniss1,2, Allison M. McKendrick1, Andrew Turpin2


1Optometry

Program #699

& Vision Sciences, 2Computing & Information Systems, The University of Melbourne, Australia

Purpose
-Establish bounds on the accuracy of visual sensitivity predictions from structural imaging that would reduce measurement error and test duration in perimetry -Establish whether current structure-function relations fall within these bounds
Figure 1 (left): a) Mean error, b) SD error and c) number of presentations for ZEST and SZEST using SM predictions 0, 10, 20, 30dB across the range of true thresholds. Data are shown for 1000 repeats per true threshold incorporating typical false positive response errors (FPR 15%, FNR 3%).

Methods
Computer simulation was used to determine the error distribution and presentation count for Structure-ZEST (SZEST), a modified ZEST algorithm using a gaussian prior centred on the threshold prediction from a hypothetical structural measure (SM). Prior PDF and termination criteria were optimised for perfect SM prediction of true threshold. SZEST was applied to the full range of true thresholds, using the full range of possible SM predictions (1000 repeats per threshold/prediction combination). The accuracy of SM predictions for which SZEST was less variable than ZEST was measured. A hypothetical structure-function relationship was then simulated (1000 repeats) using ZEST measurement error, with SM simulated to predict true threshold with just adequate precision for SZEST to outperform ZEST. Four response conditions were modelled [false positive rate, false negative rate]: [0%, 0%], [15%, 3%], [3%, 15%], [20%, 20%].

Conclusions
A perimetric procedure incorporating sensitivity predictions from structural measures (SZEST) typically makes more accurate and faster threshold estimations than ZEST when the structural measure predicts threshold within +/-10dB. This level of precision may be achievable with current clinical imaging tools.

Figure 2 (right): Mean +/-2SD error and presentation count for SZEST when accuracy of SM prediction varies. Data are shown for typical false positive observers (a) and typical false negative observers (b).

Results
SZEST performed best with PDF SD 5dB and termination at PDF SD <1.5dB. When response errors were made, SZEST was less variable than ZEST when the SM prediction was within +/-10dB of true threshold (Figure 3). With no response errors, ZEST and SZEST variability was similar. With this prediction accuracy, SZEST was faster than ZEST across all response conditions (Figure 4). The simulated structure-function relationship where SM predicted true threshold to within +/-10dB had mean R2 0.39 (95% range 0.19 to 0.56) once typical measurement error from the standard ZEST algorithm was incorporated (Figure 5).

Figure 3 (above): a) SM prediction accuracy vs. reduction in variability using SZEST (cf. ZEST). Median, IQR and range are shown for typical response error rates. Within the shaded region SZEST was less variable than ZEST for >80% of true thresholds at each prediction accuracy, median (IQR) reduction in SD 1.62dB (0.59 to 2.64). b) As (a) for damaged locations only. Within the shaded region median (IQR) SD reduction was 1.91dB (0.90 to 2.85).

Figure 4 (above): a) As Figure 3a for presentation count. Within the shaded region (SM prediction within +/-10dB of true threshold) SZEST made median (IQR) 6.3 (4.1 to 7.3) fewer presentations than ZEST. b) As (a) for damaged locations only. Within the shaded region median (IQR) 6.8 (5.7 to 7.4) fewer presentations by SZEST.

Figure 5 (right): Simulated structure-function relationships. a) A perfect underlying relation is assumed. b) Gaussian measurement noise is added such that SM predicts VF sensitivity to within +/-10dB. c) Typical ZEST measurement variability is added to simulate a clinically measured relationship.

Support: Australian Research Council, Heidelberg Engineering

jdenniss@unimelb.edu.au

You might also like