7 Summary The work presented in the current dissertation explores the potential of digital simulations with virtual reality technology, simulated prosthetic vision and deep neural networks for the optimization of cortical visual prostheses. Specifically, the presented research focuses on natural tasks for functional prototyping, computational optimization frameworks for automated optimization, and biological considerations for improving the translational value of simulation research. Chapter2describes a study with 21 sighted participants who navigated in a real-world, visually complex environment using a virtual reality (VR) setup with simulated prosthetic vision (SPV). The study tested two levels of environmental visual complexity, two image processing algorithms and six phosphene resolutions. With a simple scene representation, 26 x 26 simulated phosphenes were found sufficient to achieve >90% of the obstacle avoidance performance compared to direct camera vision. Deep learning-based contour detection resulted in equal or significantly lower mobility performance compared to edge detection. Physical removal of the surface textures in the environment resulted in improved mobility with low phosphene counts, but slightly lower performance with higher phosphene counts. These findings indicate that a functional scene processing strategy for mobility depends on a robust and balanced implementation that accounts for the hardware configuration of the implant. Chapter3investigated a reinforcement learning-based computational benchmark for more automated, task-based, evaluation of phosphene vision with simulated patients. Several experiments were performed in a virtual hallway environment that was based onChapter 2. A comparison between multiple phosphene resolutions revealed that performance increases with the number of simulated electrodes but that the performance saturates. In contrast to the experiments with human participants inChapter 2, the results of the virtual patients did not reveal an interaction effect between phosphene resolution and scene complexity. A second set of virtual experiments with different edge detection thresholds indicated that the framework can be used to evaluate optimal image processing parameters. Although the differences and similarities between sighted human participants and computational agents remain to be further explored, this study indicates that reinforcement learning can be used as a computational benchmark for the functional evaluation of prosthetic vision in a naturalistic mobility task. 111
RkJQdWJsaXNoZXIy MTk4NDMw