Enhancing artificial vision by integrating machine learning in retinal prosthetics

RWTH Publications (RWTH Aachen)

Abstract

Implantable retinal prostheses represent a promising solution for restoring partial vision but are often constrained by limited electrode resolution and non-specific neuronal stimulation, which reduces the quality of perceived visual information. To address these limitations, this dissertation proposes and evaluates machine learning-based approaches designed to enhance artificial vision. The core technical contributions are twofold. First, this work addresses the stimulation optimization problem through three novel encoding strategies with physiologically validated computational models: the reverse encoder, the end-to-end encoder, and the invertible encoder. These approaches optimize stimulation patterns by framing the problem within an encoder-decoder architecture where the encoder is trained to maximize perceptual quality. Second, a visual fixation-driven framework, inspired by biological vision, is proposed to prioritize salient image patches, thus optimizing the transmission of visual information. Subjective experiments confirmed the cross-subject consistency of fixation patterns, supporting the feasibility of this strategy. Complementing these core approaches, the dissertation also explores several auxiliary applications, including a survey of visually impaired individuals to inform development goals, a braille simulation framework for assistive reading, and retinal imaging processing pipelines for human and mouse models. Together, these contributions offer a comprehensive machine learning-based framework for improving artificial vision in retinal prosthetics, addressing both the encoding optimization challenge and the incorporation of biologically inspired visual mechanisms.

Authors 1

1. Yuli Wu corresponding Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen