Optimizing Vision Restoration: Deep Learning Approaches in Visual Cortex Decoding and Neuroprosthetic Design

This thesis aims to explore and advance the design of visual neuroprostheses, with an emphasis on neural decoding and image processing improvements. It places particular focus on the feasibility study of a device targeting the optic nerve. To address this challenging topic, the work of this thesis benefits from advancements in deep learning. This technique offers opportunities to enhance the processing of the images given as input to the prosthesis and the optimization of stimulation protocols, ultimately improving the quality of artificial vision provided to users. The work presented in this thesis specifically focuses on two key aspects: decoding visual cortex activity to investigate whether cortical activation patterns might be suitable feedback signals for a potential closed-loop device, and processing natural scenes so that relevant information for orientation and environment understanding can be extracted and fed into visual neuroprostheses. The first aspect was approached by investigating the use of convolutional neural networks (CNNs) to successfully decode visual stimuli from primary visual cortex (V1) activity in mice, enabling us to consider it as a potential source of feedback in the optic nerve stimulation context. The findings suggest that deep learning can be instrumental in designing closed-loop systems that steer stimulating electrodes based on real-time cortical responses. The second aspect was explored by exploiting deep learning to preprocess the images optimizing their quality and selecting meaningful content, and developing a geometric model to simulate retinal stimulation (RS) and optic nerve stimulation (ONS). Simulated prosthetic vision (SPV) experiments were conducted on healthy subjects, which allowed us to compare the efficacy of ONS and RS, highlighting ONS’s potential advantages for dynamic tasks requiring broad visual field awareness. This work highlights the significant impact that deep learning and geometric modeling can have in this field. These findings demonstrate the potential for these techniques to enhance the design and functionality of visual prosthetic systems. The successful application of deep learning for image preprocessing and decoding, combined with our innovative geometric model for stimulation simulations, paves the way for more effective and personalized vision restoration solutions. Looking ahead, further exploration and refinement of these methods hold promise for overcoming current challenges and achieving even greater improvements in prosthetic vision capabilities.