GIN-transformer based pairwise graph contrastive learning framework.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Resting-state functional magnetic resonance imaging (rs-fMRI) is crucial for diagnosing neuropsychiatric disorders like autism spectrum disorder (ASD) and major depressive disorder (MDD).
• Current deep learning models for rs-fMRI analysis require extensive labeled data, hindering their clinical application.

Purpose of the Study:

• To develop an unsupervised deep learning framework, GITrans-PairCL, to overcome data scarcity in rs-fMRI analysis for neuropsychiatric disorders.
• To integrate Graph Isomorphism Network (GIN) and Transformer architectures for multi-scale feature extraction from rs-fMRI data.

Main Methods:

• Proposed a GIN-Transformer-based pairwise graph contrastive learning framework (GITrans-PairCL) with Dual-modal Contrastive Learning (DCL) and Task-Driven Fine-tuning (TDF) modules.
• DCL utilizes sliding-window augmented rs-fMRI time series, employing GIN for local spatial connectivity and Transformer for global temporal dynamics.
• Cross-view contrastive learning was used for multi-scale feature extraction, followed by fine-tuning for downstream classification tasks.

Main Results:

• GITrans-PairCL demonstrated superior performance compared to traditional machine learning and deep learning baselines in automatic brain disease diagnosis.
• The model achieved high accuracy in both single-site and cross-site evaluations on public datasets.
• The framework effectively combines local and global features, reducing reliance on labeled data.

Conclusions:

• The GITrans-PairCL framework offers a promising unsupervised approach for diagnosing brain diseases using rs-fMRI data.
• This method enhances model generalization and reduces the need for extensive labeled datasets in clinical settings.
• The integration of GIN and Transformer architectures enables effective multi-scale feature learning for improved diagnostic accuracy.