gms | German Medical Science

Objective: Modern prototypes of closed-loop stimulation systems include stimulation and recording front-ends for interacting with the tissue. To analyze the extracellular activity for understanding retinal processes of the retinal tissue, further signal processing with machine learning support is necessary. These methods should also be implemented in hardware as a retinal signal processor. The implementation of these hardware accelerators have to fulfill hard constraints for using in neural implants (low space, limited power budget, low latency, high data rate processing). This abstract presents methods for enable neural signal processing in future retinal implants.

Concept and Methods: From the extracellular recordings, it is possible to classify four subtypes (ON/OFF, transient/sustained) of retinal ganglion cells (RGC) using deep learning methods. Here, the classification is done on the detected spike frames from the transient input. For transferring the models into hardware, elasticAI.creator is used for building the digital accelerators on FPGAs. In addition, it is possible to fit the impedance value from transient stimulation signals (current and voltage) to get the electrical impedance spectroscopy and the corresponding electrical model parameters. This can be used for determining the electrode quality during run-time.

Results: The 1D-CNN-based classification model for RGC subtypes consist of five CNN layers with PReLU activation function and MaxPooling and of five dense-layers with ReLU activation function. In total, it achieves a precision of 73.84% and the detection of ON/OFF RGCs achieves a precision of 81.24% after 100 epochs with 80/20 dataset splitting using Retinal Ganglion Cell Typology Database. In extracting the impedance properties from the transient electrical stimulation, a mean absolute percentage error of 9% and a difference of the tissue resistance is below 6% are achieved, compared to the reference from electrical impedance spectroscopy