Transcript Introduction to Machine Intelligence

Neuroprosthetics
Week 8
Visual Neuroprostheses
History






Brindley (Cambridge) tried a series of
experiments in the 1950s – limited success, but
opened the field
Last 15 years – lots of initial tests
Mostly animal studies – proof? Of concept
Limited human studies
First generation will be pixelated vision for the
profoundly blind (avoid guide dog?)
Mostly still speculation/experimental
Physiology of Visual Pathway - 1






Best site for an implant yet to be resolved
Different sites – different characteristics
Anatomy – www.webvision.med.utah.edu
Light falls on the retina – located at the
back surface of the eye
Photoreceptor neurons (in the retina)
convert electromagnetic (light) energy into
electrochemical signals
These are first stage retinal neurons
Physiology of Visual Pathway - 2





Output of retinal ganglion cells (last)
collected together on the optic nerve
Fibres reorganised at the optic chiasm
Majority form synapses in the lateral
geniculate nucleus (LGN) of the thalamus
LGN neurons project to cerebral cortex
Region called visual cortex
Jargon




Receptive field- type of visual stimulus that
causes neuron to respond
Visuotopic – map from visual to neural space
Visual pathway – massively parallel? signal
processing
M (large) and P (small) are two segregated
pathways thought to represent (M) where object
is and (P) what object is
Blindness





Mainly age-related degeneration, retinitis
pigmentosa (RP), accidents and cancers
Also glaucoma, diabetes but treatable
Age related leading cause (2M in USA) – mainly
loss of fine detail – central photoreceptors
degenerate
RP – inherited, affects peripheral + night vision
– leads to tunnel vision – rod photoreceptors go
Accidents + cancer more difficult as whole eye
may be lost or visual pathway affected
Prosthesis - Key Elements




Recipients must be aware that their
resultant sight will not be perfect/normal
Acceptable system must be almost invisible
Components integrated into glasses etc
First generation experimental systems may
not satisfy these criteria
Video Encoder






Mimics photoreceptors in the retina
For cortical or optic nerve based - CCD array or
photodiode array:
Conventional, cheap video cameras good for lab
exp.
For retinal based – could be integrated into the
neural interface, so reside in the plane of the
retina:
Latter has advantage of using natural
acquisition, so no robotic head movements
Spatial resolution low – limited no. of electrodes
Signal Processing






Visual system organised as a hierarchical
sequence of maps
Visuotopy: close points in space excite close
together neurons – low resolution conformal but
locally random
Prediction of light spots only ½ degree
Signal encoded into discrete signals – one for
each neural electrode
Light adapted into range of stimulus levels –
must not be affected by ambient light
Image compression + remapping for perception
Telemetry & Power








Wireless link for power and video signals in
implant
RF or light transmission – cellphone tech
Telemetry – bidirectional, circuitry informs
external electronics of power needs
Transmitter + Receiver only 1cm apart
Receiving coil implanted in eye or scalp
Frequency must be limited to avoid heat and
radiation damage to tissues
Implanted receiver small, high reliability
Low BW better but more electrodes means more
BW
Neural Stimulator







Video signals processed before stimulating
neurons
External stimulator electronics easier
Preferable for complete implant – problems
Requires on-chip memory locations
Each location dedicated to each electrode
VLSI device prototyped
Hermetic sealing of electronics difficult
Neural Interface




Same problems as with other neural
interfaces
Biological Biocompatibility
Physical Biocompatibility – implant density,
barriers, mechanical compliance, wire
tethering
Percutaneous – v - implant
Four Approaches




Subretinal
Epiretinal
Optic Nerve
Cortex
Subretinal Approach







Replication of photoreceptors – good approach
for most cases, uses remnant bipolar cells
Array of phototransducers is placed in the
subretinal space (Artificial Silicon Retina)
Each element is photodiode + electrode
Resultant voltage gradient from light source
stimulates bipolar cell dendrites
No external power or control needed
Little/no signal processing required
Presently undergoing human trials
Epiretinal Approach





Stimulating electrodes on inner retina surface –
excite remnant ganglion cells
Array of electrodes attached to inner retina
surface
Patterns of electrodes stimulated electrically
Simple, linear organisation
Still just ideas – can it be permanently attached?
Can useful signals be obtained at safe currents?
Optic Nerve Approach






Optic nerve sole visual conduit from retina to
lateral geniculate nucleus in thalamus
Only one human study – spiral cuff electrode array
with four surface electrodes
Biphasic pulses, thresholds 350 microA
Stimulus rate 8 to 10 Hz
Identify simple objects via a head mounted camera
Method not good for high resolution – MEA better ?
Cortically Based Approach









Yet to be developed – practically??
Dobelle – stimulation via electrode arrays under the
dura on visual cortex
percutaneous connector behind the ear
Each electrode connected to one of 64 pins
Subjects able to perceive points of light
Currents 1 to 10 mA (unsafe for chronic implant?)
Local electrodes alter image – nonlinearities – so
large spacings required
Recently single electrodes – 10 microA
MEA thought to be the way to go!
Final Words





Four poss sites – subretinal, epiretinal, optic
nerve and visual cortex
Passive photodiode arrays cannot produce
currents that excite retinal neurons
Stimulating electrodes must be positioned close
to neurons to excite them – electrodes must
have same dimensions
Stimulation best with highly localised current
injections – penetrating electrodes
Electrode arrays felt to be the way ahead – first
human trial was in 2002!!!!!!
Next Week

Motor Neuroprostheses