Deep Feature Extraction for Resting-State Functional MRI by Self-Supervised Learning and Application to Schizophrenia...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Functional magnetic resonance imaging (fMRI) is a key tool for understanding brain function.
• Traditional fMRI analysis often requires explicit labels or predefined regions, limiting its scope.
• Self-supervised learning offers a promising avenue for leveraging unlabeled neuroimaging data.

Purpose of the Study:

• To develop a novel self-supervised deep learning technique for functional MRI (fMRI) analysis.
• To investigate the potential of using subject identity as a supervisory signal in neural network training for fMRI.
• To explore the utility of the learned features for both subject identification and psychiatric disorder diagnosis, such as schizophrenia.

Main Methods:

• A novel self-supervised learning scheme was developed for fMRI analysis.
• A neural network was trained using functional MRI scans, with subject identity serving as the teacher signal.
• The network was trained without requiring explicit labels, relying solely on the inherent information within the fMRI data.
• Conventional methods like region of interest (ROI) pooling and principal component analysis (PCA) were used for comparison.

Main Results:

• The proposed method demonstrated that individual temporal volumes of resting-state fMRI contain sufficient information to identify subjects.
• The trained neural network learned a feature space where features clustered by subject, outperforming conventional methods.
• A simple linear classifier applied to subject-specific features achieved classification accuracy comparable to conventional functional connectivity methods for schizophrenia diagnosis.
• The extracted features captured brain functioning related to both subject identity and psychiatric disorder diagnosis.

Conclusions:

• The proposed self-supervised learning technique effectively utilizes subject identity from fMRI data.
• The learned features demonstrate potential for identifying subjects and aiding in the diagnosis of psychiatric disorders like schizophrenia.
• This deep learning approach offers a valid and powerful new design for self-supervised learning in neuroimaging analysis.