Algorithms for Epiretinal Prostheses
Collaborators: Karthik Ganesan, Nishal Shah, Lauren Grosberg, Pulkit Tandon
Advised by: E.J. Chichilnisky, Subhasish Mitra
Table of contents
An epiretinal prosthesis is a device that replaces the function of outer retinal neurons (photoreceptors) lost to disease (AMD, retinitis pigmentosa) by electrically stimulating the retinal ganglion cells (RGC) layer. In principle, it captures the visual image with a camera, processes the image in ways that mimic the function of the healthy retina, and electrically stimulates remaining RGCs to transmit visual signals to the brain that can restore vision fully or partially to people blinded by photo-receptor loss.
Project 1: Axon Bundle Detection Algorithm V1
Collaborators: Karthik Ganesan
A major challenge with epiretinal prostheses is unwanted activation of axon bundles that lie between the electrodes and target RGCs when trying to activate the RGC cell bodies (somas). Axon bundle activation was defined to be (at the time of this project) a prominent bidirectional electrical signature observed on the electrical recordings from micro-electrode arrays upon electrical stimulation. For a prosthetic device to function as intended, it is essential to identify the stimulation levels that result in axon bundle activation and hence avoid it by remaining below the stimulation threshold.
For that purpose, we developed a completely automated algorithm based on recorded electrical waveforms. The algorithm involved the following two steps: (i) graph creation & partitioning. We first partitioned the electrode array by creating a graph using the recorded waveforms amplitudes. On partitioning this graph using spectral clustering methods, we were able to recover spatially disjoint clusters that qualitatively matched the physical axon bundle layouts. (ii) bidirectional propagation identification. We then tracked the electrical activity’s center of mass along space and time to identify the threshold for the classic bifurcation of signal of axon bundle activation. Our automated algorithm yielded activation thresholds that matched ground truths identified by manual analysis with with 0.93 correlation.
Our work won the best poster award amongst >200 participants with a cash prize of $1000 at the Stanford BioX Seed Grant Symposium and was published with neurobiological experimental results and validation led by Lauren Grosberg in the Journal of Neurophysiology. This paper was awarded the distinction in scholarship by the American Physiological Society (APSselect certificate).
Figure 1: Axon Bundle Detection Algorithm V1
Project 2: Constrained Sequential Decision Making & Closed-loop Algorithms for Epiretinal Prosthesis
Collaborators: Nishal Shah
A complete epiretinal prosthetic digital ecosystem includes implantable components with stringent power constraints (\(< 1 mW / mm^{2}\)) & size limitations (\(~10 mm^{2}\))and custom SoCs (on body system-on-chip) for on-chip run-time computation & wireless communication with strict real-time latency constraints and limited power. Our project was on developing and implementing algorithms that can operate within the tight constraints yet accurately determine spatio-temporal patterns for electrically stimulating the RGCs when given an input video from the camera to produce the best visual perception to the patient.
Direct naive implementation of algorithms which generate accurate stim-patterns computationally (a hard task by itself) will exceed our resource budget by a large amount. Instead, we adopted joint cross-disciplinary, neuroscience-driven co-optimization of architectural and algorithmic techniques to meet this challenge. Our preliminary results indicated almost an order of magnitude improvement in speed and power by employing neuroscience insights, hardware-software techniques such as reduced precision arithmetic, caching, sparsity, implicit parallelism, etc.
Furthermore, we modelled the stim-pattern problem as a Markov Decision Process (MDP) with States given by perception, Actions by the electrical-stimulation dictionary, Transition Probabilities measured as the change in perception on stimulation and the Reward signal was the improvement in similarity to the target perception.
Our joint work led us to being one the top 30 finalist teams from over 200 at the Qualcomm Innovation Fellowship competition.
Figure 2: Epiretinal Prosthetic System
Project 3: Axon Bundle Detection Algorithm V2
Collaborators: Pulkit Tandon
As the prosthetic system evolved, the definition of what constitutes problematic axon bundle activation in the context of an epiretinal prosthesis changed from “emergence of a bidirectional spike pattern at a certain stimulation amplitude” to “axonal stimulation of RGCs with unknown soma and receptive field locations, typically outside the electrode array”. Consequently, the problem changed to detection of the lowest current amplitude at which the activity in any off-array cell is observed.
The new axon bundle detection algorithm uses two main ideas: electrically-evoked spikes show lower variance than spontaneously occuring spikes, and with an increase in stimulation amplitude, we expect to see a monotonic increase in spike probability.