Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Synthetic Data for Discriminating Serotonergic Neurons using Convolutional Neural Networks

Daniele Corradetti, Alessandro Bernardi, Renato Corradetti

TL;DR

This work tackles the challenge of identifying serotonergic neurons, including atypical cells, from electrophysiological recordings where conventional visual criteria are insufficient. It introduces a synthetic data augmentation pipeline that smooths real spike waveforms and fuses them with diverse real noise masks to expand training data for CNN classifiers. A CNN ensemble trained on these synthetic samples achieves strong generalization to non-homogeneous data (e.g., accuracy ~0.94, TPR ~0.962, TNR ~0.888, AUC ~0.926), and a practical case is provided with open-source code. The approach provides a scalable, real-time capable tool for serotonergic neuron identification, with future work pointing to GAN-based augmentation and broader species/dataset applicability.

Abstract

Serotonergic neurons in the raphe nuclei exhibit diverse electrophysiological properties and functional roles, yet conventional identification methods rely on restrictive criteria that likely overlook atypical serotonergic cells. The use of convolutional neural network (CNN) for comprehensive classification of both typical and atypical serotonergic neurons is an interesting one, but the key challenge is often given by the limited experimental data available for training. This study presents a procedure for synthetic data generation that combines smoothed spike waveforms with heterogeneous noise masks from real recordings. This approach expanded the training set while mitigating overfitting of background noise signatures. CNN models trained on the augmented dataset achieved high accuracy (96.2% true positive rate, 88.8% true negative rate) on non-homogeneous test data collected under different experimental conditions than the training, validation and testing data.

Synthetic Data for Discriminating Serotonergic Neurons using Convolutional Neural Networks

TL;DR

This work tackles the challenge of identifying serotonergic neurons, including atypical cells, from electrophysiological recordings where conventional visual criteria are insufficient. It introduces a synthetic data augmentation pipeline that smooths real spike waveforms and fuses them with diverse real noise masks to expand training data for CNN classifiers. A CNN ensemble trained on these synthetic samples achieves strong generalization to non-homogeneous data (e.g., accuracy ~0.94, TPR ~0.962, TNR ~0.888, AUC ~0.926), and a practical case is provided with open-source code. The approach provides a scalable, real-time capable tool for serotonergic neuron identification, with future work pointing to GAN-based augmentation and broader species/dataset applicability.

Abstract

Serotonergic neurons in the raphe nuclei exhibit diverse electrophysiological properties and functional roles, yet conventional identification methods rely on restrictive criteria that likely overlook atypical serotonergic cells. The use of convolutional neural network (CNN) for comprehensive classification of both typical and atypical serotonergic neurons is an interesting one, but the key challenge is often given by the limited experimental data available for training. This study presents a procedure for synthetic data generation that combines smoothed spike waveforms with heterogeneous noise masks from real recordings. This approach expanded the training set while mitigating overfitting of background noise signatures. CNN models trained on the augmented dataset achieved high accuracy (96.2% true positive rate, 88.8% true negative rate) on non-homogeneous test data collected under different experimental conditions than the training, validation and testing data.
Paper Structure (12 sections, 2 equations, 4 figures, 1 table)

This paper contains 12 sections, 2 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Limitations of Visual Identification Methods
  3. Problems with the Deep Learning Methods
  4. Generation of Synthetic Data
  5. A Practical Case
  6. Original Training Data:
  7. Non-homogenous Data:
  8. Synthetic Data:
  9. The Architecture of the Neural Network:
  10. Results on Test Data
  11. Results on Non-Homogenous Data
  12. Conlusion and Discussion

Figures (4)

  • Figure 1: Appearance of extracellularly recorded action potentials according to the recording arrangement. A. Spikes from a serotonergic neuron recorded in voltage recording configuration, typically used in vivo or in vitro with sharp microelectrodes. Left trace shows two spikes on a slow recording timebase. Note the predominance of background noise over the spikes in the recorded segment. Right trace shows one of the spikes at faster timebase. Note the rapid positive upstroke followed by a negative phase comprising two subsequent downstrokes. B. Appearance of the same signals as obtained with the loose-seal patch-clamp current recording configuration used in the present work.
  • Figure 2: Example on how single events were isolated and selected. The image depicts the recording of the serotonergic cell A131031#170 and the 4 ms event of triggered at point 1068617, i.e. at 26.715 sec.
  • Figure 3: Example of 4 synthetic spikes generated by the event triggered at 1068617, i.e. 26.715 sec, of the serotonergic cell A131031#170. Top trace: the original recording of the event. The panels report four spike obtained by processing the original trace with different noise masks (see methods).
  • Figure 4: Examples of noise masks collected from the recordings of cell A160125#009 (on top) and A150701#086 (on bottom).