Spike Train Scalograms (STS): a Deep Learning Classification Pipeline for Neuronal Cell Types.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Classifying neuronal cell types is essential for understanding cerebral cortex circuitry.
• Traditional Machine Learning methods rely on manually engineered features, potentially missing complex spike train patterns.

Purpose of the Study:

• To introduce a novel Deep Learning (DL) pipeline for accurate neuronal cell type classification.
• To leverage Spike Train Scalograms (STS) and Continuous Wavelet Transform (CWT) for analyzing electrophysiological (EP) data.
• To compare the DL approach against traditional Machine Learning (ML) baselines.

Main Methods:

• Developed a DL pipeline integrating CWT scalograms with pre-trained Convolutional Neural Network (CNN) architectures.
• Applied the pipeline to patch-clamp EP recordings from 5,590 murine cortical neurons.
• Utilized InceptionV3 CNN architecture and employed explainability analyses (saliency maps, SHAP).

Main Results:

• Achieved high classification accuracy for neuronal cell types (Pvalb, Sst, Vip/Lamp5, Excitatory), with balanced accuracy of 90.53% and weighted F1-Score of 90.03%.
• The STS pipeline effectively handled class imbalances.
• Explainability analysis showed strong agreement between the DL model, ML baseline, and known biological characteristics.

Conclusions:

• The novel STS-based DL pipeline offers a highly accurate method for neuronal cell type classification.
• This approach effectively captures complex spike train dynamics using spectral analysis and DL.
• The method requires significantly less data (two raw sweeps) compared to traditional methods.