Skip to content Skip to navigation

MEA-based classification of retinal ganglion cells for bionic vision.

Hamed Shabani 1, Mahdi Sadeghi 1, Zohreh Hosseinzadeh 1,2, Eberhart Zrenner 1, Daniel Rathbun 1,3


1 Institute for Ophthalmic Research, Eberhard Karls University, Tübingen, Germany
2 Department of Molecular and Cellular Mechanisms of Neurodegeneration, University of Leipzig, Germany
3 Department of Ophthalmology, Henry Ford Health System, Detroit, MI, USA


Purpose: Although there has been significant progress in developing retina implants during last two decades, due to the inability to selectively stimulate different Retina Ganglion Cell (RGC) types, visual perception for retina implant patients remains limited. We hypothesize that different types of RGCs can be selectively activated by deriving stimuli from their different electrical input filters. The input filters of cells are extracted from their response to electrical noise stimulation using the Spike Triggered Averaging (STA) method. To begin testing this hypothesis, we first classify RGC types using a set of visual stimuli and then examine the properties of each cell type’s electrical input filters.


Methods: In this study we used the data recorded from nine dark adapted retinas of seven adult wild type mice. A 60 channel microelectrode array in contact with the ganglion cell side of the retina was used to record the spiking neural activity of RGCs. The visual stimulation set was adapted from Baden et al. (Nature 2016), including moving bars, contrast and temporal frequency chirps, blue-green color flashes, and spatiotemporal white noise. In order to extract electrical input filters, a sequence of filtered and interpolated Gaussian white noise voltage steps was used. Similar to Baden et al. we used sparse principle component analysis (sPCA) to extract response features to the visual stimuli.
After projecting data into a lower-dimensional space, we assigned each neuron to one of the 75 clusters reported by Baden et al., by finding the highest correlation between a neuron’s response and the clustered response data provided by Baden et al.
Results: We recorded visual responses from 426 RGCs. These responses mapped onto about half of the previously described clusters. Despite convolving our spike trains with a filter to create pseudo-calcium traces for correlation with the previous dataset, many of our responses were significantly more transient than previously reported. ON and OFF cells had different electrical input filters as we have previously reported.


Discussion: Adaptation of the Baden et al. methodology for spike trains instead of calcium recordings was partially successful. For better classification results, new cluster definitions should be derived from a large spike train data set. Electrical input filters do appear to vary with RGC type, but more precise cluster definitions are needed to refine this result.


Financial disclosure: None

Go Back