Neural activity shaping in visual prostheses with deep learning.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Artificial Intelligence

Background:

• Retinal prostheses face limitations in visual perception due to current spread from adjacent electrodes, reducing spatial resolution with unipolar stimulation.
• Neural activity shaping (NAS) using simultaneous multipolar stimulation can improve control over neural activation patterns by attenuating excessive excitation spread.

Purpose of the Study:

• To propose and validate a model-free artificial neural network (ANN) approach for neural activity shaping (NAS) in retinal prostheses.
• To develop an efficient and personalized method for retinal stimulation to improve visual prosthesis outcomes.

Main Methods:

• A two-stage ANN system was developed: a measurement predictor network (MPN) trained on implant data to predict retinal response, and a stimulus generator network trained on natural images.
• The stimulus generator network utilized the MPN to learn the inverse model for determining efficient multipolar stimulus patterns.
• Validation was performed in silico using a realistic retinal response model.

Main Results:

• The ANN-based NAS approach demonstrated sharper retinal activations compared to conventional unipolar stimulation.
• The ANN strategy achieved results equivalent to analytical model inversion (AMI) but was model-agnostic and computationally more efficient (three orders of magnitude).

Conclusions:

• The novel ANN-based protocol enables efficient and personalized retinal stimulation.
• This method holds potential for significantly improving the visual experience and quality of life for retinal prosthesis users.