A Model Of Information Processing In The Retina

Objective:Visual prosthesis could produce phosphenes by electrically stimulating the structures on visual pathway, which helps the blind to get visual information around. As an essential part in the visual prosthesis research, the retinal model of information processing can reproduce the biological function of the retina, transforms the visual image to spike trains. Through theoretical analysis of the signal transfer function and information flow in the retina, a bionic model is established to solve the problems about what information should be inputted to the visual center, how input information could effectively be transmitted. We hope the model to be helpful for the design and development of implantable visual microelectronic system.Methods:Based on state-of-the-art retinal physiological mechanism, a two-layered visual information model is brought forward. It includes the outer plexiform layer for information extraction and the inner plexiform layer for information encoding. The model is simulated on the Matlab Platform in this paper.1. Outer plexiform layer for visual information extraction. The receptive fields of retinal neurons display spatial opposition and temporal delay between the center and surround regions. In our model, the functions of both retinal center regions and surround regions correspond to a linear spatiotemporal filter. The filter's spatial component is a Gaussian kernel, and its time component is the convolution of several exponential kernels. Finally, bipolar cells layer integrates both center output signals and surround output signals, which are always antagonistic.2. Inner plexiform layer for visual information coding. Static nonlinear rectification is very common feature in neural modeling and in retinal models. A simplified synaptic transmission function was used to model the signal shaping in the amacrine cells layer. The model can generate the spike trains with a simple homogeneous Poisson model, in which a stochastic refractory period is randomly chosen after each spike. 3. A graphical user interface was developed with the GUIDE, an integrated development environment provided by Matlab.Results:1. We can get the contour edge of the image through the model of outer plexiform layer. Using the model of inner plexiform layer, we can get the spike trains, which convey the visual information to the brain.2. For a static grayscale image through the spatial-temporal filtering, compared with the spatial Difference of Gaussian model, not only can our model retain the contour edge of the image, but also have large areas of contrast on the input image.3. When the relative weight (ω=0.95) of the surround signal to the center signal, we get clear contour edge image, which is consistent with the other experimental results.4. Sigmoid static nonlinear rectification function allow different input luminosity with varying scaling factors. Also we prove its pertinence by reproducing the experiment measurements from single ganglion cells.5. The main interface shows us the name and features of system, and the sub-interface includes four parts:images loading, spatial-temporal filtering, static non-linear rectification, non-uniform radial sampling and Poisson spike trains generation.Conclusions:1. The model of retinal information processing is established on the Matlab 7.7 Platform in this paper. The information processing function of retina can be effectively simulated by the two-layered model.2. The spatial-temporal filter model of the outer plexiform layer can achieve the visual elements of low-level extraction capabilities. Comparing with the traditional Gaussian differential model, not only it can keep the high spatial frequency components of the visual image, but also can retain the low Spatial frequency components. 3. Sigmoid static nonlinear rectification function ensures the fire rate of ganglion cells within reasonable range.4. Graphical user interface layout is simple and reasonable. Graphical user interface combines dynamic simulation with parameters setting, making the model faster and more convenient in application.