Title: Enhancing Epiretinal Implant Perception Through Model-Based Deep Reinforcement Learning and In Silico Stimulation

Abstract

Objective: Conditions like retinitis pigmentosa and age-related macular degeneration lead to the deterioration of the photoreceptor layer. To counteract this vision loss, one therapeutic strategy involves electrically activating remaining retinal ganglion cells using microelectrode arrays, such as those found in epiretinal implants. However, these implants typically produce anisotropic, elongated visual percepts (phosphenes) that align with the axon fascicles of adjacent retinal ganglion cells. Previous studies suggest that by mapping these axon fascicles and either deactivating specific electrodes or reducing stimulation currents to avoid triggering them, it is possible to achieve more isotropic, pixel-like shapes. The primary goal of avoiding axon fascicle stimulation is to eliminate the "brushstroke" effect, thereby creating a simpler, more uniform set of pixel-like visual cues.

Approach: This research introduces a novel method using both isotropic and anisotropic shapes to render recognizable images for a virtual patient within the "rlretina" reinforcement learning environment. This environment frames the challenge as a stroke-based rendering task, where the system must effectively utilize brushstroke-like inputs.

Main Results: We developed a deep reinforcement learning agent capable of combining isotropic and anisotropic shapes to construct images. The study evaluates various error-based and perception-based metrics to determine the most effective reward signals for training the agent. Utilizing a psychophysically validated axon map model, the agent is trained via a model-based data generation approach to simulate how different virtual patients perceive the rendered images. Our findings indicate that the trained agent produces significantly more intelligible images than naive methods across various virtual patient simulations.

Significance: This study presents a fresh perspective on epiretinal stimulation, marking a foundational step toward enhancing visual acuity in artificial vision systems by leveraging anisotropic phosphenes.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC