CGAN-rIRN: a data-augmented deep learning approach to accurate classification of mental tasks for a fNIRS-based...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Functional near-infrared spectroscopy (fNIRS) offers robust brain activity monitoring for brain-computer interfaces (BCIs).
• Traditional machine learning classifiers (MLCs) for fNIRS struggle with manual feature engineering, limiting accuracy.
• Deep learning classifiers (DLCs) are promising but require extensive data and computational resources, and often overlook fNIRS signal characteristics.

Purpose of the Study:

• To develop a novel, data-augmented deep learning classifier (DLC) for accurate mental task classification in fNIRS-BCI systems.
• To address the limitations of existing DLCs by incorporating temporal and spatial properties of fNIRS signals.
• To improve the accuracy of classifying voluntary mental tasks for enhanced BCI control.

Main Methods:

• Proposed a hybrid deep learning approach combining a convolution-based conditional generative adversarial network (CGAN) for data augmentation with a revised Inception-ResNet (rIRN) based DLC.
• Utilized CGAN to generate synthetic fNIRS signals, augmenting the training dataset with class-specific data.
• Designed the rIRN architecture with serial spatial and temporal feature extraction modules (FEMs) to capture multi-scale signal properties.

Main Results:

• The proposed CGAN-rIRN approach significantly improved single-trial classification accuracy for mental arithmetic and mental singing tasks.
• Demonstrated superior performance compared to traditional MLCs and commonly used DLCs in fNIRS-BCI paradigms.
• The data augmentation and classification strategy effectively enhanced the classification of brain activation patterns.

Conclusions:

• The novel data-augmented DLC (CGAN-rIRN) effectively classifies mental tasks in fNIRS-BCI, overcoming limitations of existing methods.
• This hybrid deep learning approach offers a promising, data-driven solution for improving the performance of volitional control fNIRS-BCI systems.
• The study highlights the potential of tailored deep learning architectures and data augmentation for advancing neurotechnology.