Microstimulation Encoding Techniques for Thalamic Visual Prostheses

Abstract

Visual prostheses have recently shown success in partially restoring vision to the blind by electrically stimulating the visual pathway. Techniques that target the Lateral Geniculate Nucleus (LGN) of the thalamus represent one promising type of visual prostheses. However, challenges in tuning such prostheses include understanding how visual information is encoded in the firing of LGN neurons and designing the appropriate electrical stimulation that would produce the same neural activity evoked from natural visual stimulation. In this thesis, we examined the Artificial Neural Networks (ANN) and deep Convolutional Neural Networks (CNN) techniques in the task of developing computational neural coding models for the LGN. First, neural encoding models were developed to predict the LGN neural responses to visual and electrical stimulations. Second, an electrical decoding model was then developed that could generate electrical stimulation patterns that would be used to induce visual sensation. In the visual prosthesis process, obtaining the required electrical stimulus can be achieved by using the visual encoding model to predict the response of LGN neurons to a given visual scene, then inversely use the electrical decoding model to obtain the required electrical stimulus. Our models are verified using real neural experimental data recorded from rats in vivo. For the visual encoding models, the best mean correlation between the predicted and actual firing rates of 0.75 was achieved. Our results were compared to other visual encoding techniques such as Kalman filter and Generalized Linear Model (GLM) achieving 0.27 and 0.64 mean correlation respectively. The performance of the electrical encoding and decoding models was also measured achieving 0.53 and 0.44 mean correlation, respectively. Our results demonstrate the potential of the developed models to be utilized in tuning electrical stimulation patterns for thalamic visual prostheses.