Improving the Resolution of Prosthetic Vision through Stimulus Parameter Optimization

Abstract

Retinal prostheses have helped improve vision in patients blinded by photoreceptor degeneration. These electronic implants operate by electrically stimulating the remaining retinal cells. While retinal prosthesis users report improvements in light perception and performing basic visual tasks, their ability to perceive shapes and letters remains limited. Investigating stimulation strategies to reduce perceptual thresholds and create focal, non-overlapping phosphenes will increase the resolution of retinal prostheses and improve the overall patient outcomes. In this thesis I explore two main strategies for electrode-specific optimization of stimulation parameters: 1) a novel pulse paradigm for threshold reduction, and 2) an automated closed-loop method for adjusting stimulation parameters to create a focal retinal activation area. I combined human subject testing and computational modeling to investigate the effect of waveform asymmetry on perception shapes and thresholds with epiretinal stimulation. Threshold measurement and phosphene shape analysis was performed on four Argus II users. A computational model of a retinal ganglion cell (RGC) was created in the NEURON simulation environment to allow for a more thorough parameter testing and to gain insight into the biophysical mechanisms. Our human subject results suggest that asymmetric waveforms could increase perception probabilities compared to a standard symmetric pulse, and this effect can be intensified by addition of an interphase gap (IPG). Our computational model predicts that the most effective pulse for threshold reduction is asymmetric anodic-first stimulation with small duration ratios (≤ 5) and long IPGs (≥ 2 ms). Phosphene shape analysis revealed no significant difference in percept elongation with different pulse types. Average phosphene area was larger with asymmetric anodic-first stimulation compared to other pulse types. Prosthetic vision quality is highly dependent on the capability to precisely activate target neurons and avoid off-target activation. However, studies show elongated and inconsistent responses to single electrode stimulation, indicating unintended stimulation of off-target neurons and electrode-specific activation patterns. While tuning stimulation parameters can transform the spatial RGC activity, a manual search for optimal parameters can be time consuming and tiring for patients. I developed a process for automatic optimization of stimulation parameters in silico, which involved training neural networks for quantifying the relationship between pulse parameters and spatial response descriptors, and a closed-loop algorithm to search for optimal parameters. Using this process, I was able to guide the parameter search effectively and converge to an optimal response within a few iterations. Finally, I presented a process for automatic optimization of stimulation parameters in vitro using calcium imaging in mouse retina. This process involved training neural networks at each iteration based on a few images, using an interior point algorithm to find the optimal parameters, and classifying the resulting calcium images with a CNN trained on previous data. Our results indicate that we can converge to optimal stimulation parameters that create focal RGC activity by sampling less than 1/3 of the parameter space. This approach can shorten the exploration time significantly compared to a manual search, especially when the parameter space is large. Findings of this project could lead to the development of a clinically applicable system for electrode-specific optimization of stimulation protocol, improving the overall outcome of artificial vision.