Separability of Adjacent Neurons Recorded with a CMOS-Multi-Transistor-Array

Abstract

Event Abstract Back to Event Separability of Adjacent Neurons Recorded with a CMOS-Multi-Transistor-Array Christian Leibig1, 2, 3*, Armin Lambacher4, Thomas Wachtler1, 5 and Günther Zeck2 1 Ludwig-Maximilians-Universität München, Department Biology II, Germany 2 Natural and Medical Sciences Institute at the University of Tübingen, Germany 3 University of Tübingen, Graduate School for Neural Information Processing, Germany 4 Max-Planck-Institute of Biochemistry, Membrane- and Neurophysics, Germany 5 Bernstein Center for Computational Neuroscience Munich, Germany Background/Aims In order to understand the functionality of neural networks, reliable parallel spike trains must be recorded from neuronal ensembles [1, 2]. Silicon-based multi-transistor arrays ('Neurochips') are large enough to map the activity of hundreds to thousands of neurons within an area of 1 mm^2. In order to retrieve reliable parallel spike trains, action potentials (APs) have to be properly assigned to their corresponding units. The signal separation of adjacent neurons in neural tissue or cell culture represents a serious challenge, as their electrical coupling areas on the sensor array overlap. In this work we aim to determine whether this problem can be solved based on extracellular action potential waveforms together with their positional information. Methods A potential algorithm for assigning APs to corresponding units was presented in [3]. Here, we test its capabilities by using synthetic data with known positions, signal shapes and amplitudes of the extracellular signal. The synthetic data were generated from recordings of four retinal ganglion cells (RGCs). Measurements were done with a CMOS-multi-transistor array (size 1 mm^2, pitch 7.4 µm, sampling rate 78 kHz). For each RGC several hundred threshold crossings were averaged and down sampled to either 11.5 or 23 kHz respectively. Gaussian noise was added with appropriate amplitude to obtain a signal-to-noise ratio of 11 [4]. Thus a template was obtained for each RGC. Finally, synthetic data sets were generated by combining template pairs at different distances from each other (0, 7.4, 14.8 or 22.2 µm) and fed to the algorithm described in [3]. The quality of spike sorting was quantified by an error rate defined as the sum of false positive and false negative assignments with respect to the known ground truth. This procedure was carried out for two sampling rates (11.5 and 23 kHz). Results We compared spikes from pairs of different neuronal templates with centres of electrical coupling areas either aligned on each other or separated by multiples of 7.4 µm. For both temporal sampling rates, spikes could be reliably separated down to the smallest separation distance (0 µm). Conclusion/Summary For given conditions (signal-to-noise ratio, noise distribution and templates) our results provide parameter settings for which spiking activity can be reliably assigned to the corresponding source. Acknowledgements Supported by BMBF grant 0312038.