{"title": "Dopaminergic Neuromodulation Brings a Dynamical Plasticity to the Retina", "book": "Advances in Neural Information Processing Systems", "page_first": 559, "page_last": 565, "abstract": null, "full_text": "Dopaminergic Neuromodulation Brings a \n\nDynamical Plasticity to the Retina \n\nEric Boussard \n\nJean-Fran~ois Vibert \n\nB3E,  INSERM  U263 \n\nFaculte de  medecine Saint-Antoine \n\n27  rue  Chaligny \n\n75571  Paris cedex  12 \n\nAbstract \n\nThe fovea of a mammal retina was simulated with its detailed bio(cid:173)\nlogical  properties  to  study  the local  preprocessing  of images.  The \ndirect  visual  pathway (photoreceptors,  bipolar and  ganglion cells) \nand the  horizontal units,  as  well  as the  D-amacrine cells  were sim(cid:173)\nulated.  The  computer  program simulated the  analog  non-spiking \ntransmission between photoreceptor and bipolar cells,  and between \nbipolar and ganglion cells,  as well as the gap-junctions between hor(cid:173)\nizontal cells,  and the release  of dopamine by  D-amacrine cells  and \nits  diffusion  in  the  extra-cellular  space.  A  64 x 64  photoreceptors \nretina,  containing  16,448  units,  was  carried  out.  This  retina  dis(cid:173)\nplayed  contour  extraction  with  a  Mach  effect,  and  adaptation  to \nbrightness.  The simulation showed that the dopaminergic amacrine \ncells  were  necessary  to  ensure  adaptation to  local brightness. \n\n1 \n\nINTRODUCTION \n\nThe  retina is  the first  stage  in  visual  information processing.  One  of its  functions \nis  to  compress  the information received  from  the environment by  removing spatial \nand  temporal  redundancies  that  occur  in  the  light  input  signal.  Modelling  and \ncomputer simulations present  an efficient means to investigate and characterize the \nphysiological  mechanisms that  underlie  such  a  complex  process.  In  fact,  filtering \ndepends  on the  quality of the input image  (van Hateren,  1992): \n\n559 \n\n\f560 \n\nBoussard and Vibert \n\nl.High mean light intensity (high  signal  to noise  ratio).  A  high-pass  filter  en(cid:173)\nhances  the edges  (contour extraction)  and the temporal changes of the input. \n\n2.Low mean  light intensity (low  signal  to  noise  ratio).  The sensitivity of high(cid:173)\npass filters  to noise  makes them inefficient  in this case.  A  low-pass filter,  averaging \nthe signal over several  receptors,  is  required  to extract  the relevant information. \n\nThere are three aspects in the filtering adaptivity displayed by the retina:  adaptivity \nto  i)  the  global  spatial  changes  in  the  image,  ii)  the  local  spatial  changes  in  the \nimage, iii)  the temporal changes in the image.  We will focus  on the second feature. \nA  biologically  plausible  mammalian retina was  modelled and simulated to explore \nthe  local  preprocessing  of the images.  A first  model  (Bedfer  &  Vibert,  1992),  that \ndid  not  take into account  the  dopamine neuromodulation, reproduced  some of the \nbehaviors  found  in  the  living  retina,  like  a  progressive  decrease  of ganglion  cells' \nfiring  rate  in  response  to  a  constant  image  presented  to  photoreceptors,  reversed \npost-image,  and  optic  illusion  (Hermann  grid).  The  model,  however,  displayed  a \npoor local adaptivity.  It could not give both a  good contrast rendering and a  Mach \neffect.  The  Mach  effect  is  a  psychophysical  law  that  is  characterized  by  an  edge \nenhancement  (Ratliff,  1965).  The  retina  network  produces  a  double  lighter  and \ndarker  contour  from  the  frontier  line  between  two  areas  of different  brightness  in \nthe stimulus.  This phenomenon is indispensable for  contour extraction.  This paper \nwill  first  present  the  conditions  in  which  high-pass  filtering  and  low-pass  filtering \noccur  exclusively  in  the  retina  model.  These  results  are  then  compared  to  those \nobtained  with  a  model  that includes  dopamine neuromodulation,  thus  illustrating \nthe role played by  dopamine in local adaptivity (Besharse  & Iuvone,  1992). \n\n2  METHODS \n\nThe retina is  an unusual neural structure:  i)  the photoreceptors  respond to light by \nan  hyperpolarization,  ii)  signal  transmission from  photoreceptors  to  bipolar  units \ndoes  not  involve  spikes,  neurotransmitter  release  at  these  synapses  is  a  continu(cid:173)\nous  function  of the  membrane  potential  (Buser  &  Imbert,  1987).  Only  ganglion \ncells  generate  spikes.  Furthermore,  horizontal  cells  are  connected  by  dopamine \ndependent  gap-junctions.  Dopamine is  an  ubiquitous neurotransmitter and  neuro(cid:173)\nmodulator in  the  central  nervous system.  In  the  visual pathway,  dopamine  affects \nseveral types of retinal neurons  (Witkovsky & Dearry, 1992).  Dopamine is  released \nby stimulated D-amacrine and  interplexiform cells.  It diffuses  in  the extra-cellular \nspace,  and produces:  cone  shortening  and rod elongation,  reduced  permeability of \ngap-junctions, increased conductance of glutamate-induced current among horizon(cid:173)\ntal cells,  increased  conductance  of the  cone-to- horizontal cell  synapse,  and  retro(cid:173)\ninhibition on  D-amacrine cells  (Djamgoz & Wagner,  1992).  Our  model focused  on \nthe adaptive filtering mechanism in the fovea that enables  the retina to simultane(cid:173)\nously  perform  both  high-pass  and  low-pass  filtering.  Therefore,  dopamine  action \non  gap-junction  between  horizontal  cells  and  the  retro-inhibition  on  D-amacrine \ncells  was  the  only  dopamine effect  implemented  (fig.  1).  Our  model included  the \nthree  neuron  types  of the  direct  pathway - photoreceptors,  bipolar  and  ganglion \nunits - as  well  as  two types of the indirect  pathway - the horizontal and dopamin(cid:173)\nergic  amacrine  cells.  Only  the  On  pathway of a  mammal fovea  was  studied  here. \nEach neuron type has  been modelled with its own anatomical and electrophysiolog-\n\n\fDopaminergic Neuromodulation Brings a Dynamical Plasticity to the Retina \n\n561 \n\n~ Excitation  . .  lnhibition \n\n.11 II'  Gap-junction  ~ ~fea:!ne \n\nFigure  1:  The  dopaminergic amacrine units in  the  modelled retina. \n\nThe  connections  of an  On  center  pathway  in  the  simulated  retina.  Photo:  Pho(cid:173)\ntoreceptors.  Horiz:  Horizontal  units.  Bip:  Bipolar  units.  Gang:  Ganglion  Units. \nDA:  Dopaminergic Amacrine unit.  DA  units  are  stimulated by  many  bipolar units. \nWith  an  enough  excitation,  they  can  release  dopamine  in  the  extracellular  space. \nThis  released  dopamine  goes  to  modulates the  conductance  value  of horizontal gap(cid:173)\njunctions. \n\n\f562 \n\nBoussard and Vibert \n\nical  properties  (Wiissle  & Boycott,  1991)(Lewick  & Dvorak,  1986).  The  temporal \nevolution of the  membrane potential of each unit  can  be  recorded. \n\n3  RESULTS \n\nA  64x64  photoreceptors  retina was  constructed  as  a  noisy  hexagonal  frame  where \nphotoreceptors,  bipolar  and  ganglion  units  were  connected  to  their  nearest  neigh(cid:173)\nbours.  Horizontal units were connected to their 18 nearest photoreceptors and bipo(cid:173)\nlar  units,  with  a  number of synaptic boutons  decreasing  as  a  function  of distance. \nThey did  not retroact  on the nearest  photoreceptor.  This horizontal layer architec(cid:173)\nture  produces  lateral inhibition.  Each modelled D-amacrine unit  was  connected  to \nabout  fifty  bipolar  units.  The  diffusion  of released  dopamine  in  the  extra-cellular \nspace  was  simulated.  The  modelled  retina  consisted  of  16,448  units  and  862,720 \nsynapses. \n\nAt  each  simulation,  the  photoreceptors  layer  was  stimulated  by  an  input  image. \nStimulations were given as a 256x256 pixel image presented to the simulated 64x64 \nphotoreceptor retina.  Since the localization of photoreceptors  was not regular, each \nreceptor  received  the  input from  16  pixels  on  the  average.  The output  image  was \nreconstructed using the ganglion units response.  For each of the 4096 ganglion units \nthe spike frequency  was measured during a given time (according to the experiment) \nand coded  in  a  grey  level for  the given  unit  retinotopic  position.  Thus, each simu(cid:173)\nlation  produced  an  image of the  retina output.  This  output  image was  compared \nto  the input image. \n\nThe input image  (stimulus) consisted here  of one  white  disk  on a  dark background. \nThe  results  presented,  in fig.  2,  were  obtained  after  750  ms of stationary stimula(cid:173)\ntions.  The stimuli were  here  a  white  disk  on  a  black  background.  The inputs were \nstationary  to  avoid  temporal  effects  owing  to  evolving  inputs.  Output  images  of \nstationary  inputs,  however,  vanished  after  1000  ms.  The  time was  limited  to  750 \nms  to optimize the quality of the output image. \n\nBiological datas available on the conductance value suggest  that in the mammalian \nretina the conductance does not remain constant and undergoes  a dynamical tuning \ndepending on the local brightness [?].  This provides a range of possible values for the \nconductance.  The  behavior  of the  model  was  tested  for  values  within  this  range. \nDifferent  values  lead  to  different  network  behaviors.  Three  types  of results  were \nobtained from  the simulations : \n\nl.Without dopamine action, the conductance values were fixed for  all gap-junctions \nto 1O- 6S (fig.  2-A). The output image rendered well the contrast in the input image, \nbut did  not display the  Mach effect  (low-pass filtering). \n\n2.Without dopamine action, the conductance values were fixed  for  all gap-junctions \nto  1O- 10S (fig.  2-B).  The low conductance  value allowed a pronounced  Mach effect, \nbut the contrast  in  the output image was strongly diminished  (high-pass filtering). \nThis  contrast  appears  like  an  average  of  the  two  brightness.  Only  the  contour \ndelimited by Mach effect  allows the  disk  to  be distinguished. \n3.With dopamine, the conductance values were initially set to  1O- 7S (fig.  2-C) . The \noutput  image  displayed  both  the  contrast  rendering  and  the  Mach  effect  (locally \n\n\fDopaminergic Neuromodulation Brings a Dynamical Plasticity to the Retina \n\n563 \n\nA \n\n40 \n\n32 \n\n24 \n16 \n\n8 \n\n\u00b00~---1-1--~22~-3~4---4~5--~56 \n\n22 \n23 \n24 \n26 \n27 \n28 \n29 \n30 \n32 \n33 \n34 \n\n'\n\nB \n\n40 \n32 ..-----,. \n24 \n16 \n\n8 \n\n0~~1~2--~24~~3~6--~4=8--~60 \n\nc \n\nLJV'\\....I'-_I \n\n65 \n52 \n39 \n\n26 \n13 \n\n0~~12~~2~4--~3~7--4~9~~61 \n\nFigure  2:  ~ontour extraction  (Mach  effect)  according  to  gap-junctions  conduc(cid:173)\ntances. \n\nOn  the  left,  results  obtained  after  750  ms  of stimulation  for  an  zmage  of a  white \ndisk  on  a black  background.  On  the  right,  sections through  the  corresponding image. \nA bscissa:  spike  count;  Ordinates:  geographic  position  of the  unit,  from  the  left  side \nto  the  middle  of the  left  panel.  A:  without  dopamine  (fixed  Ggap  =  10- 6 S).  B: \nwithout  dopamine  (fixed Ggap  = lO-lOS).  C:  with  dopamine  release  (starting Ggap \n= 1O- 7 S) .  A  gives  a good  contrast  rendering,  but  no  Mach  effect.  B  gives  a  Mach \neffect,  but  there  is  an  averaging  between  darker and  lighter areas.  C,  with  dopa min(cid:173)\nergic  neuromodulation,  gives  both  a Mach  effect  and  a good  contrast  rendering. \n\n\f564 \n\nBoussard and Vibert \n\nadaptive filtering). \n\n4  DISCUSSION \n\nThese  results  show  that  the  conductance  cannot  be  fixed  at  a  single  value  for  all \nthe  gap-junctions.  If the  conductance  value is  high  (fig.  2-A),  the  model acts  like \na  low-pass filter.  A  good  contrast  rendering  was  obtained,  but there  was  no  Mach \neffect.  If the  conductance  value is  low  (fig.  2-B),  the  model  becomes  a  high-pass \nfilter.  A  Mach  effect  was  obtained,  but  the  contrast  in  the  post-retinal image was \ndramatically deteriorated:  an  undesirable  averaging of the  brightness  between  the \ndarker and the more illuminated areas  appeared.  Therefore in this model the Mach \neffect  was only obtained at the expense of the contrast.  A mammalian retina is able \nto  perform  both  contrast  rendering  and  contour  extraction  functions  together.  It \nworks  like  an  adaptive filter.  To  obtain a  similar result,  it  is  necessary  to  have  a \nvariable  communication between  horizontal units.  The simulated retina  needs  low \ngap-junctions  conductance  in  the  high  light  intensity  areas  and  high  conductance \nin  the  low  light  intensity  areas.  The  conductance  of each  gap-junction  must  be \ntuned  according  to  the  local  stimulation.  The  model  used  to  obtain  the  fig.  2-C \ntakes into account the dopamine release by the D-amacrine cells.  Here,  the network \nperforms  the  two  antagonist functions  of filtering.  Dopamine  provides  our  model \nwith  the  capacity  to  have a  biological  behaviour.  What  is  the  action  of dopamine \non  network?  Dopamine is  released  by D-amacrine units.  Then,  it  diffuses  from  its \nrelease  point into the extra-cellular space among the neurons,  reaches gap-junctions \nand decreases  their  conductance value.  Thus the conductance  modulation depends \nin time and in intensity on the distance between gap-junction and D-amacrine unit. \nIn  addition, this  action is  transient. \n\n5  CONCLUSION \n\nThanks  to  dopamine  neuromodulation,  the  network  is  able  to  subdivise  itself \ninto  several  subnetworks,  each  having  the  appropriate  gap-junction  conductance. \nEach subnetwork is  thus adapted for  a  better  processing  of the  external  stimulus. \nDopamine neuromodulation is  a chemically addressed  system, it acts more diffusely \nand  more  slowly  than  transmission  through  the  axo-synaptic  connection  system. \nTherefore  neuromodulation adds a  dynamical plasticity to the network. \n\nReferences \n\nG.  Bedfer & J .-F. Vibert.  (1992) Image preprocessing in simulated biologicalretina. \nProc.  14th  Ann.  Conf.  IEEE EMBS 1570-157l. \nJ.  Besharse  & P.  Iuvone.  (1992).  Is  dopamine a  light-adaptive or  a  dark-adaptive \nmodulator in retina?  NeuroChemistry International 20:193-199. \nP.  Buser  & M.  Imbert.  (1987)  Vision.  Paris:  Hermann. \nM.  Djamgoz & H.-J.  Wagner.  (1992)  Localization and function  of dopamine in the \nadult vertebrate  retina.  NeuroChemistry InternationaI20:139-19l. \n\nL.  Dowling.  (1986)  Dopamine:  a  retinal neuromodulator?  Trends  In  NeuroSciences \n\n\fDopaminergic Neuromodulation Brings a Dynamical Plasticity to the Retina \n\n565 \n\n9:236-240. \nW . Levick & D.  Dvorak.  (1986)  The retina - from molecule to network .  Trends  In \nNeuroSciences 9:181-185. \n\nF.  Ratliff.  (1965)  Mach  bands:  quantitative  studies  on  neural network in the  retina. \nHolden-Day. \n\nJ . H.  van Hateren.  (1992)  Real and optimal images in early vision.  Nature 360:68-\n70. \nH.  Wassle  &  B.  B.  Boycott. \nretina.  Physiological Reviews 71(2):447-479. \nP . Witkovsky &  A.  Dearry.  (1992)  Functional roles  of dopamine in  the vertebrate \nretina.  Retinal Research 11:247-292. \n\n(1991)  Functional  architecture  of  the  mammalian \n\n\f", "award": [], "sourceid": 768, "authors": [{"given_name": "Eric", "family_name": "Boussard", "institution": null}, {"given_name": "Jean-Fran\u00e7ois", "family_name": "Vibert", "institution": null}]}