Purchase this article with an account.
Raphael Sznitman, Serife Kucur, Pablo Marquez-Neila, Mathias Abegg, Sebastian Wolf; A Reinforcement Learning method for Perimetry Testing. Invest. Ophthalmol. Vis. Sci. 2019;60(9):4382.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Perimetry is a non-invasive clinical psychometric examination used for diagnosing ophthalmic and neurological conditions and suffers from being a slow examination method. Maintaining high levels of concentration during this time is exhausting for patients and negatively affects the acquired visual field (VF). Our hypothesis is that using only a subset of test locations in a VF pattern, VFs can be accurately estimated in less time.
We designed a novel perimetry testing strategy named Patient Attentive Sequential Strategy (PASS), based on Deep Learning and reinforcement learning. PASS daptively selects what locations to query based on the patient's answers to previous queries in order to reconstruct the complete visual field as accurately as possible, and then separately estimates the visual field from the sparse observations. PASS accordingly has two interacting parts: Policy and Reconstruction functions. Policy function was modeled using a neural network (NN) which assigns probabilities to each location to be selected the next. We used a NN or a Least Squares approach that we refer PASS+RNet and PASS+LSTSQ, respectively. We tested our approach on data acquired at the Rotterdam Eye Institute (Netherlands) that includes 5108 VFs from 22 healthy and 139 glaucomatous patients using the 24-2 pattern by the Humphrey Visual Field Analyzer II. Performance was measured using the Mean Square Error (MSE) between the estimated and full threshold acquired VFs.
When testing for 8 locations, PASS yielded an MSE=27.85 with 26.98 stimuli presentations on average, while Tendency Oriented Perimetry (TOP) strategy that led to MSE=32.35 with 54 presentations. When testing 36 locations, PASS had an MSE=13.82 with 108.07 stimuli presentations, compared to Dynamic strategy with MSE=15.17 and 156.31 presentations. Qualitatively, PASS selects locations dynamically and dependent on the patient eye.
Our experiments show that PASS outperforms state-of-the-art methods on two different datasets, leading to more accurate reconstructions while reducing between 30% and 70% the duration of the patient examination. Further validation on is necessary, but PASS could provide for a faster and more accurate VF testing method for glaucoma management.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
Quanlitative comparison between our proposed method (PASS+RNET/LSTSQ) and SORS. For each row, we show how different methods estimate a visual field using increasing amounts of stimuli presentations.
This PDF is available to Subscribers Only