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

Abstract

Classifying neuronal cell types is crucial for understanding the intricate circuitry of the cerebral cortex, which comprises diverse specialized neurons essential for brain function. Traditional Machine Learning (ML) approaches rely on manually engineered electrophysiological (EP) features, often overlooking subtle and complex patterns within spike train data. This study introduces a novel Spike Train Scalograms (STS)-based Deep Learning (DL) pipeline that integrates Continuous Wavelet Transform (CWT) scalograms with pre-trained Convolutional Neural Network (CNN) architectures to classify neuronal cell types with high accuracy. Utilizing patch-clamp EP recordings from 5,590 murine cortical neurons, the pipeline transforms spike trains into time-frequency representations via CWT, capturing both transient and sustained signal characteristics. These scalograms are then processed by fine-tuned CNN architectures, including InceptionV3, which achieved a balanced accuracy and weighted F1-Score of 90.53% and 90.03%, respectively. The STS pipeline effectively distinguishes between major neuronal types such as Pvalb, Sst, Vip/Lamp5, and Excitatory neurons, even in the presence of class imbalances. Moreover, an explainability analysis using saliency maps and SHAP revealed high correspondence between the DL approach, the ML baseline and biological knowledge of these neuronal types. The results demonstrate that, by combining an advanced spectral analysis with DL techniques, neurons can be classified with high accuracy, employing only two raw sweeps rather then the full stimulation range required by shallow approaches.