Localizing neuronal somata from Multi-Electrode Array in-vivo recordings using deep learning.

Area of Science:

• Computational Neuroscience
• Neuroimaging Techniques
• Machine Learning in Biology

Background:

• High-density Multi-Electrode Arrays (MEAs) enable sophisticated in-vivo neural recordings.
• Advances in computational neural modeling allow simulations of neurons with detailed morphology.
• Accurate neuron localization is crucial for interpreting neural signal spatiotemporal information.

Purpose of the Study:

• To develop and validate a deep learning approach for determining the 3D positions of neuronal somata.
• To leverage simulated neural activity data for robust neuron localization.
• To assess the method's accuracy across diverse neural morphologies and alignments.

Main Methods:

• Utilized multi-compartment models of 13 distinct layer 5 rat neocortical neuron morphologies.
• Simulated neural recordings from neurons placed at random locations and orientations relative to MEAs.
• Employed Convolutional Neural Networks (CNNs) trained on sodium trough and repolarisation peak features from MEA data.

Main Results:

• The deep learning model accurately predicted neuronal somata 3D positions with low error rates.
• The approach demonstrated robustness across various neural morphologies and neuron-MEA alignments.
• Forward modeling combined with machine learning provided highly accurate localization results.

Conclusions:

• Deep learning effectively extracts 3D neuronal location information from simulated MEA recordings.
• The method offers a powerful tool for precise neuron localization in complex neural circuits.
• This approach enhances the interpretation of neural data from advanced in-vivo recording technologies.