3.4. Discussion 3 35 Figure 3.10: Performance for different phosphene resolutions (c.q. the number of implanted electrodes). Panel A: results from thebaseline study with sighted subjects (n=21) using SPV (adapted from: de Ruyter van Steveninck et al., 2022b). Panel B-C: Test performance of the virtual implant users (n=3). B: the number of box collisions. C: the obtained reward. 3.4. Discussion This study investigated an RL-based computational framework for the task-based evaluation of prosthetic vision. We tested the performance of a virtual patient that learns via an RL-framework to navigate in a virtual hallway using SPV. The study setup is based on a prior SPV study with sighted subjects (de Ruyter van Steveninck et al., 2022b), andthe results are explicitly compared. 3.4.1. Primary outcomes Baseline - virtual navigation with RL The baseline results presented inFigures 3.6to3.8indicate that our model successfully converged to a visually-guided navigation strategy using Q-learning. The results from perturbation analysis indicate that the model makes use of sensible visual feedback such as the presence of boxes directly in front of the camera, or the unobstructed view on objects in the distance. After training, the model reaches similar performance (collisionfree navigation) compared to sighted human subjects. Edge detection threshold The found inverted-U-shape for the mobility performance over different edge detection thresholds indicate that our RL-model is affected by suboptimal image processing parameters. An inherent challenge for many image processing strategies is that they require fine-tuning. This is not unique for basic image processing algorithms such as edge detection. Often, more complex or intelligent image processing strategies such as limited depth rendering, object detection, or contour prediction (de Ruyter van Steveninck et al., 2022b; Vergnieux et al., 2017; Wang et al., 2022b) also require fine-tuning of confidence or thresholding parameters. The example case results of our experiments illustrate how our RL-benchmark could be implemented to search optimal image processing parameters. Phosphene resolution The positive general relationship between mobility performance and the phosphene resolution is in line with the results in thebaseline study byde Ruyter van Steveninck
RkJQdWJsaXNoZXIy MTk4NDMw