gms | German Medical Science

Objective: Actual prototypes of retinal implants are performing an open-loop stimulation. The effectiveness of this stimulation protocol decreases during run-time due to changes within the retinal tissue. With integrating a recording front-end for sensing the extracellular retinal activity, the bi-directional link allows to implement a retinal signal processor (RSP) for understanding retinal processes. To enable a closed-loop stimulation, an encoder strategy is necessary for linking the retinal activity to the stimulation parameters. This abstract presents a concept for closed-loop stimulation with deep learning techniques in computational manner.

Concept and Methods: We propose a closed-loop processing pipeline based on the feedback of the retinal ganglion cell activation ratio (RGC-AR). An encoder gets the target and the actual RGC-AR as input and adjusts the stimulation parameters. We assessed an artificial neural network as encoder architecture using a simple static retina simulator. To adjust for dynamic change, we propose a Recurrent Neural Network (RNN) with long short-term memory (LSTM) cells as encoder. In the retina simulator, the output values are superimposed with a linear drift and a sine waveform.

Results: Initial results indicate that the encoder in the static case adjusts well to the retina simulator with a mean absolute error of ca. 1%. The dynamic approach demonstrates that LSTM-based RNN architecture is suitable in principle for adapting to the retina's dynamic behavior. The closed-loop stimulation approach is viable in a computational context, but further research is needed (i) to determine the capability of RNNs to compensate for retinal dynamics and (ii) to optimize the RSP and encoder for hardware implementation using in retinal implants. In the future, we will move forward to a more realistic retina simulator based on real experimental data.

Acknowledgment: This work was supported by the German Research Foundation grant 424556709.