5 72 5. Towards biologically plausible phosphene simulation 16 epochs are visualized inFigure5.8. Note that the model successfully converged to a sparse phosphene encoding that selectively represents the object boundaries. Enlarged and inverted visualizations can be found inFiguresS6andS7. Small electrode counts and interactions The previously described experiments simulate implant designs with many electrodes. As a supplementary experiment, we verified that the end-to-end model can also be trained with smaller phosphene counts. The simulator was initialized with a random subset of 60 electrodes from a 10×10 array with electrode spacing 0.4 mm, matching the location and spacing of array no. IV in Figure 5.1. An image dataset was used with white characters on a black background. We tested three conditions to address potential interaction effects between neighboring electrodes. Besides a baseline training, we evaluated training the model with an additional loss component to avoid unexpected interactions by discouraging simultaneous activation of neighboring electrode pairs. This co-stimulation loss component is defined as Lcostim= 1 n nX i=1 mX j=1µ Ii · Ij 1+∥pj −pi ∥ 2¶, (5.14) for stimulation currents Ii and Ij , and ∥pj −pi ∥ 2 the squared distance between the electrodes in mm. In a final training condition, instead of avoiding interactions using a loss component, we explicitly included an interaction model in the phosphene simulation. An electrode coactivation interaction was implemented, where current of active electrodes ‘leaks’ to neighbouring activated electrodes based on their distance. For each active electrode pair i and j we added the coactivation current Icoact = mX j=1µ Ij 1+100∥pj −pi ∥ 2¶ (5.15) to the stimulation current Ii used in the simulation. The results are visualized inFiguresS8andS9. The found encoding strategy resulted in distinct letter shapes. The letters are, however, poorly recognizable which is unsurprising with the minimal electrode resolution. The costimulation loss resulted in a lower percentage of active neighbouring electrodes (at a distance < 1mm). The electrode coactivation resulted in a higher percentage of active neighbouring electrodes compared to the baseline, suggesting that the encoder learns to make use of the leak current. 5.4. Discussion The aim of this study is to present a biologically plausible phosphene simulator, which takes realistic ranges of stimulation parameters, and generates a phenomenologically accurate representation of phosphene vision using differentiable functions. In order to achieve this, we have modeled and incorporated an extensive body of work regarding the psychophysics of phosphene perception. From the results presented in section Section 5.3.1, we observe that our simulator is able to produce phosphene percepts that match the descriptions of phosphene vision that were gathered in basic and clinical visual
RkJQdWJsaXNoZXIy MTk4NDMw