Implementation of Figure–Background Separation in a Learning Retina Implant

Abstract

Purpose: : To study the functional properties of a figure–background separation system (FB) for a learning epiretinal retina implant in conjunction with a corresponding simulator of the human central visual system.

Methods: : A: A figure–background separation system (FB) for a learning retina implant (Eckmiller et al., ARVO 2006) was integrated into a novel retina encoder (RE*) and simulated on a modern PC System. The FB module was designed to mimic two typical functions of the central visual system: 1) figure–background separation and 2) oculomotor mode selection in order to elicit visual ‘Gestalt’ perception in retinally blind subjects by means of epiretinal stimulation with an array of about 10 x 10 electrodes under realistic visual scene conditions. Visual patterns typically consisted of a stationary component (P_S) and a moving component (P_M) as input of RE* with spatio–temporal (ST) filters representing receptive field (RF) properties of individually tunable P– or M–type ganglion cells. B: Figure–background separation was implemented by combining the two mechanisms of: a. sequential monitoring of ST filter responses to temporal– vs. spatial input changes and b. response transfer of identified P_S– vs. P_M pixels into respective memory locations. C: Oculomotor mode selection decided about whether P_S– or P_M was chosen as figure. The selection e.g. of the smooth pursuit mode was implemented by subsequently providing the ST filter array only with the P_M memory content and thus suppressing the stationary background pixels of P_S. D: Retinal information processing was implemented by 10 x 10 ST filters with tunable spatial and temporal parameters to simulate typical P– or M–ganglion cell properties each feeding into one electrode channel. The photosensor layer as input for optical patterns P_S– or P_M was implemented as a hexagonal array with about 100 x 100 photosensor pixels. E: To mimic the human central visual system of future blind subjects with learning retina implants, an inverter module at the output of RE* was implemented. F: Alternative learning algorithms were developed for dialog–based tuning of RE*.

Results: : (1) Both FB functions: figure–background separation and oculomotor mode selection could be completely integrated into RE*. (2) Figure–background separation could be achieved within a few movement steps. (3) Implementation of smooth pursuit was achieved electronically without physically moving parts.

Conclusions: : FB as part of RE* in a learning retina implant provides future implant wearing subjects with the choice of fixation or smooth pursuit, which is essential to transform realistic, superimposed patterns into visual percepts.

Keywords: pattern vision • retinal connections, networks, circuitry • computational modeling