fNIRS-Driven Depression Recognition Based on Cross-Modal Data Augmentation.

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Early depression diagnosis is crucial for recovery but relies on subjective clinical assessments.
• Functional near-infrared spectroscopy (fNIRS) combined with deep learning shows promise for objective depression diagnosis.
• Limited fNIRS datasets hinder the application of data-hungry deep learning models.

Purpose of the Study:

• To develop an fNIRS-driven architecture for depression recognition that overcomes data limitations.
• To enhance depression diagnosis accuracy using cross-modal data augmentation techniques.
• To validate the proposed method for both binary depression diagnosis and multi-class severity recognition.

Main Methods:

• Proposed the fNIRS-driven cross-modal data augmentation (fCMDA) architecture.
• Converted fNIRS data into pseudo-sequence activation images.
• Implemented time-domain augmentation (time warping, time masking) and a stimulation task-driven pseudo-sequence method.
• Utilized deep classification networks with an imbalance loss function.

Main Results:

• Achieved high accuracy in two-class depression diagnosis (0.905).
• Demonstrated strong performance in five-class depression severity recognition (0.889).
• The fCMDA architecture proved effective in recognizing depression with limited fNIRS data.

Conclusions:

• The fCMDA architecture offers a novel and effective solution for depression recognition using fNIRS.
• Cross-modal data augmentation significantly improves model performance with limited datasets.
• This approach facilitates more accessible and objective depression assessment.