Performance of simulated visual fields using structure-derived prior information

Abstract

Purpose : In this preliminary study, we evaluated the performance of using structure-derived visual field priors (S-priors) for simulated visual fields (VFs).

Methods : Retrospective data from 1399 subjects (single eyes) from a Singapore population study were used. HFA™ II-i (ZEISS, Dublin, CA) SITA Standard 24-2 VFs and CIRRUS™ HD-OCT (ZEISS, Dublin, CA) 200x200 Optic Disc cubes were analyzed. 70% of eyes were randomly chosen and the data used to train regressors to predict the VF: i) a random forest (RF) using the 256-point circumpapillary retinal nerve fiber layer (RNFL) data and age; ii) a simplified mixed-scale dense convolutional neural net (CNN) [Pelt et al. PNAS 2018; 115(12)] using the RNFL thickness map. The remaining 30% of eyes were used as a test set to predict S-priors and as true input fields to a VF simulator.

A simulator was developed that implemented a Bayesian ZEST using a bi-modal starting probability distribution (PDF), as described previously [Chong et al. OPO 2015; 35(2)], centering the normal mode on age normal values derived from 118 eyes described previously [Flanagan et al. IOVS 2016; 57(12)].

ZESTs using a uni-modal PDF for custom priors centered on both types of S-priors were also simulated (ZEST-RF, ZEST-CNN). Slopes of frequency of seeing responses were modeled as previously described [Henson et al. IOVS 2000; 41(2)]. False answer rates were set to 0%, 5%, and 20% as 3 responder types.

Performance between simulated and true VFs was evaluated by observing mean absolute error (MAE) and total number of stimulus questions. Two one-sided, paired t-tests (α=0.05) for inter-strategy equivalence versus ZEST were performed using limits of equivalence of ±0.5 dB for MAE and ±5% for total questions.

Results : Mean VF mean deviations were -1.8±2.4 dB and -2.7±2.7 dB for training and test sets, respectively (p<0.001). The models performed better in higher vs. lower thresholds (Fig 1). However, overall MAEs for ZEST-RF and ZEST-CNN were equivalent to ZEST (see Table 1). Total questions were reduced 16-19% with S-priors.

Conclusions : The findings suggest that even models with limited data that predict VF priors from OCT data can reduce the duration of a VF exam in this population with comparable overall error. With more data representing a clinical population and more refined models, performance may be further improved.

This is a 2020 Imaging in the Eye Conference abstract.

 

Fig. 1. Overview of S-prior models

 

Table 1. Summary of MAE and total stimulus questions