Signal2Image Modules in Deep Neural Networks for EEG Classification.

Area of Science:

• Biomedical Engineering
• Computer Science
• Machine Learning

Background:

• Deep learning excels in computer vision using image data.
• Biomedical field has abundant physiological signals for diagnosis.
• Effective utilization of signals for deep neural network training is an open research question.

Purpose of the Study:

• To define and evaluate Signal2Image (S2I) modules for converting physiological signals into image-like representations.
• To compare the performance of various S2Is with established deep learning models for signal classification.
• To investigate the efficacy of trainable versus non-trainable S2I approaches.

Main Methods:

• Defined Signal2Image (S2I) as modules converting signals (e.g., Electroencephalography - EEG) to image formats.
• Compared four S2Is ('signal as image', spectrogram, 1-layer CNN, 2-layer CNN) with multiple deep learning 'base models' (LeNet, AlexNet, VGGnet, ResNet, DenseNet).
• Evaluated accuracy and time performance, including depth-wise and 1D CNN variations.

Main Results:

• The one-layer Convolutional Neural Network (CNN) S2I outperformed non-trainable S2Is in eleven out of fifteen tested models for EEG signal classification.
• Empirical evidence supports the effectiveness of trainable S2Is over non-trainable ones.
• Visual comparisons of S2I outputs were presented.

Conclusions:

• Trainable Signal2Image (S2I) modules, particularly a one-layer CNN, offer a promising approach for deep learning on physiological signals.
• This methodology enhances the applicability of image-based deep learning models in the biomedical domain.
• Further research can explore diverse S2I architectures and signal types for improved diagnostic and predictive capabilities.