Purchase this article with an account.
Yudong Tao, Mohamed Khodeiry, Rui Ma, Karam Alawa, Ximena Mendoza, Xiangxiang Liu, Mei-Ling Shyu, Richard K Lee; Deep reinforcement learning for optimized visual field analysis. Invest. Ophthalmol. Vis. Sci. 2021;62(8):1008.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Existing visual field testing algorithms rely on manually crafted rules to determine the contrast sensitivity of a light stimulus at each step as well as the stopping criterion, which is suboptimal in terms of testing accuracy and speed. We developed and tested a deep reinforcement learning (DRL)-based visual field testing algorithm that achieves greater performance gain in terms of testing accuracy, speed, and robustness than the Zippy Estimation by Sequential Testing (ZEST) algorithm.
Our algorithm formulates the procedure of white-on-white static automated perimetry (SAP) as a reinforcement learning problem, with the perimeter modeled as an intelligent agent that could interact with the patient. At each step, the agent can choose to present a stimulus with a specific intensity to the patient or activate a stop signal to end the test and predict the sensitivity. By training the agent on our visual field dataset with 275 normal and 433 glaucomatous visual fields, a new rule is learned by the agent. This rule can guide the agent to perform a series of optimized actions in each test, which leads to the highest testing accuracy with the fewest number of stimuli presentations.
The proposed algorithm was evaluated on both synthetic datasets and our patient visual field dataset under different prior distributions. Experimental results show that our algorithm is 9.4% faster and 5.6% more accurate than ZEST (p = 0.025). Meanwhile, our algorithm can generalize better than ZEST when the distribution of testing data is different from the initial distribution.
Our results demonstrate that our DRL-based visual field testing algorithm performs better than ZEST. By maintaining higher accuracy with reduced testing time, our algorithm increases clinic efficiency, improves patient satisfaction, and improves testing results with less patient fatigue. Moreover, the improved generalization ability our algorithm ensures test reliability and robustness of our algorithm.
This is a 2021 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only