Deep Learning Enhanced Label-Free Action Potential Detection Using Plasmonic-Based Electrochemical Impedance...

Area of Science:

• Neuroscience
• Biophysics
• Signal Processing

Background:

• Accurate measurement of neuronal electrical activity, like action potentials, demands high spatial and temporal resolution.
• Existing tools struggle to meet these demands, limiting detailed cellular analysis.
• Plasmonic-based electrochemical impedance microscopy (P-EIM) offers label-free, subcellular resolution mapping of action potentials but requires extensive signal averaging.

Purpose of the Study:

• To develop advanced signal processing techniques to reduce the number of averaged cycles needed for action potential detection in P-EIM.
• To improve the signal-to-noise ratio and overall usability of P-EIM for real-time neuronal activity monitoring.
• To demonstrate the efficacy of deep learning methods in analyzing P-EIM signals.

Main Methods:

• Utilized advanced signal processing techniques on P-EIM extracted signals.
• Applied matched filtering to detect action potential signals with reduced averaging.
• Employed a Long Short-Term Memory (LSTM) recurrent neural network for single-cycle action potential detection.

Main Results:

• Matched filtering successfully detected action potentials using as few as five averaged cycles.
• The LSTM recurrent neural network achieved a satisfactory Area Under the Receiver Operating Characteristic curve (AUC) of 0.855.
• Demonstrated successful detection of single-cycle stimulated action potentials using deep learning.

Conclusions:

• Advanced signal processing, particularly deep learning, significantly enhances the detection capabilities of P-EIM.
• The developed methods reduce the need for extensive signal averaging, overcoming a major limitation of P-EIM.
• This advancement improves the practical usability of P-EIM for mapping neuronal electrical signals with high resolution.