Development of interactive algorithms to infer cellular retinal functional properties on a closed-loop CMOS-MEA platform

The retina is a major sensory processing circuit for humans and as the overall nervous system, it is a complex and hierarchical large network of cells consisting of different types of neurons. These cells communicate with each other through biochemical and bioelectrical signals that, at the circuit level, lead to the encoding of different features of the visual inputs. To understand this process many studies based on electrophysiological recordings have been conducted on single-cells or a restricted pool of cells. Only with the recent technological developments, as the new generations of high-density microelectrodes arrays (HD-MEA) that exploit CMOS (Complementary Metal-Oxide-Semiconductor) technology, we have reached the capability of monitoring the light-evoked responses of several thousands of retinal ganglion cells (RGCs) in individual retinas. This study aims at advancing these experimental capabilities by contributing in the development of a closed-loop platform for investigating cellular retinal functional properties. Originally, it will allow to study these cellular properties by stimulating the retina according to its internal state associated with the spontaneous electrical activity. Indeed, even though HD-MEAs provide an accurate description of the neuronal activity, the implementation of closed-loop platforms that can exploit their spatial-spatiotemporal resolution remains very challenging due to the demanding computations required for handling the large recorded data stream at millisecond range latencies. A recently proposed hardware architecture that exploits the FPGA/CPU resources of a Xilinx ZedBoard Zynq-7000 was introduced to perform data pre-processing tasks, such as signal conditioning, filtering and spike detection, with a maximum latency of 2 ms. In this thesis, I have extended the capabilities of this closed-loop platform, first by implementing a pseudo-real-time algorithm that removes redundant information in the recorded spiking activity induced by the tight spacing among adjacent electrodes of HD-MEA. This algorithm consists in a clustering procedure that identifies spatial and temporal correlated spiking units. This approach has been extensively tested and validated in HD-MEA recordings of ex vivo mouse retina. Next, I have implemented a second closed-loop algorithm that controls visual stimuli to infer the major functional properties of retinal ganglion cells. In detail, based on the spiking response to white and black stimuli, the algorithm classifies the units in ON- or OFF-type retinal ganglion cells. Finally, in the last part of my study, I evaluated a potential approach for estimating the retinal ganglion cells receptive field size in real-time. State-of-the-art approaches allowing for an accurate and precise estimation of the receptive fields are very time-consuming. On the contrary, the approach that I have developed aims at minimizing the experimental time and the amount of data required for estimating the receptive fields of retinal ganglion cells. All in all, the interactive algorithms that I have designed, implemented and validated in this work against experimental data provide a new suite of tools to characterize the retinal circuit and to investigate how it implements visual information processing. In perspective, these algorithms can be extended to other experimental preparations and high-density recording devices, such as high-density implantable probes for in vivo studies.