Non-stationary neural signal to image conversion framework for image-based deep learning algorithms.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Deep learning models typically require 2D image data.
• Physiological signals are often recorded as 1D time-series data.
• Existing methods for converting 1D signals to 2D images may be time-consuming or inefficient.

Purpose of the Study:

• To develop a time-efficient framework for converting 1D physiological signals into 2D image representations.
• To enable the use of image-based deep learning models for analyzing physiological signals.
• To evaluate the performance of the proposed framework in classifying neural spikes and EEG data.

Main Methods:

• A novel preprocessing framework rasterizes 1D non-stationary physiological signals into 2D images using Bresenham's line algorithm with O(n) time complexity.
• A modified 2D Convolutional Neural Network (2D CNN) was employed for classification tasks.
• The framework was validated on two public datasets, including simulated neural recordings with varying Signal-to-Noise Ratios (SNR) and EEG seizure/non-seizure data.

Main Results:

• The 2D CNN achieved high accuracy in classifying three types of neural spikes across different SNRs (e.g., 99.69% at SNR 0.5, 91.98% at SNR 2.0).
• The framework successfully classified EEG epileptic seizure and non-seizure data with 97.52% accuracy, outperforming other algorithms.
• The proposed method demonstrated robustness and effectiveness in signal classification.

Conclusions:

• The developed time-efficient framework successfully converts 1D physiological signals into 2D images for deep learning.
• This approach significantly enhances the classification accuracy of neural spikes and EEG data.
• The methodology holds potential for application to other biomedical signals like Electrocardiograph (EKG) and Electromyography (EMG).