Abstract
Purpose :
One of the difficulties in static white-on-white automated perimetry (SAP) are fatigue effects, which make the patient responses less reliable with time. Thus, perimetry should be as fast as possible while still providing accurate estimation of the visual field. Our hypothesis is that the threshold of a visual field location can be inferred using thresholds of its neighbor locations. Following this hypothesis, we developed a novel SAP strategy that aims to reduce test time compared to conventional methods without compromising accuracy.
Methods :
We use a conditional random field (CRF) to comprehensively model each visual field location and its relationship with its neighbors. The model is trained on a dataset of 4863 visual fields of 139 patients. While individual locations are tested using zippy estimation by sequential testing (ZEST), the visual field model is constantly updated yielding new estimates for untested locations. Four locations are measured and the next location to measure is chosen by entropy and spatial threshold gradient. Locations with entropy or gradient magnitude smaller than a specified value are not tested but directly inferred from the CRF. We evaluate our approach, which we call Spatial Entropy Pursuit (SEP), in simulation using a dataset of 128 visual fields from patients with various eye diseases. The results are compared to common methods i.e. the dynamic test strategy (DTS) and tendency oriented perimetry (TOP) in terms of error and test time. For each measurement, the error was quantified by the root-mean square deviation and the test time was quantified by the number of stimulus presentations.
Results :
SEP allowed for a visual field acquisition in 128 ± 32 steps i.e. stimulus presentations and yielded similar accuracy levels as the DTS. For the same visual field acquisitions the DTS needed 154 ± 12 steps. This represents a speed up of 17%.
Conclusions :
The results demonstrate that a more data driven approach can lead to significant speed up in perimetry. Even tough the model was only trained on glaucomatous visual fields it allowed for a speed up on a broad range of eye diseases. Larger speed ups are likely if SEP is trained on data of a broader range of diseases or if the model is specific to the disease under observation. However, to achieve this, much larger datasets for a range of eye diseases are needed.
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.