Unsupervised Joint Domain Adaptation for Decoding Brain Cognitive States From tfMRI Images

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Brain decoding research aims to understand human cognitive function using neuroimaging data.
• Current methods require extensive high-quality labeled data, increasing costs and expert annotation time.
• Cross-individual decoding is challenging due to data distribution inconsistencies from individual variations and equipment differences.

Purpose of the Study:

• To propose a Join Domain Adaptive Decoding (JDAD) framework for unsupervised decoding of cognitive states from neuroimaging data.
• To address the limitations of existing brain decoding methods, including high costs and cross-individual performance issues.
• To enable more accurate and efficient brain decoding by leveraging unlabeled data.

Main Methods:

• Utilized volumetric feature extraction from task-based functional Magnetic Resonance Imaging (tfMRI) data.
• Developed a novel objective loss function incorporating a joint distribution regularizer.
• The regularizer constrains the conditional and marginal probability distributions between labeled and unlabeled samples.

Main Results:

• The JDAD framework demonstrated superior performance compared to prevalent methods on the Human Connectome Project (HCP) S1200 dataset.
• Achieved significant improvements in decoding accuracy, particularly for fine-grained tasks (11.5%-21.6% increase).
• Visualized learned 3D features using Grad-CAM, linking them to brain functional regions for group-level cognitive task analysis.

Conclusions:

• JDAD offers a robust solution for unsupervised brain decoding, overcoming limitations of traditional supervised approaches.
• The framework effectively handles individual variations and reduces the need for extensive expert-labeled neuroimaging data.
• Provides a novel pathway for understanding brain cortex functions related to specific cognitive tasks at a group level.