Combining biophysical modeling and deep learning for multielectrode array neuron localization and classification.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Neural circuits comprise diverse neuron types, making it challenging to analyze individual contributions to neural activity.
• Current methods for classifying and localizing neurons from extracellular recordings require substantial human input, leading to slow, biased, and unreliable results.

Purpose of the Study:

• To develop an automated and objective framework for classifying and localizing neurons using deep learning and detailed neuron models.
• To improve the accuracy and reliability of neural circuit mapping by overcoming limitations of current manual approaches.

Main Methods:

• Utilized convolutional neural networks trained on simulated extracellular action potentials from biophysically detailed neuron models.
• Developed a framework leveraging spatiotemporal signal profiles recorded by multielectrode arrays for automated analysis.

Main Results:

• Achieved highly accurate prediction of neuron positions, outperforming current state-of-the-art methods.
• Demonstrated very high accuracy in classifying neurons as excitatory or inhibitory, surpassing traditional clustering techniques.
• Showed potential for differentiating subtypes of excitatory and inhibitory neurons.

Conclusions:

• The proposed framework offers an automated, objective, and more accurate approach to localizing and classifying neurons from extracellular recordings.
• This advancement facilitates more precise and reliable mapping of neural circuits, contributing to a deeper understanding of neural function.