Decoding Brain States From fMRI Signals by Using Unsupervised Domain Adaptation.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning advances medical image analysis, particularly decoding brain states from fMRI.
• Cross-subject decoding using fMRI is hindered by individual differences and varying acquisition parameters, leading to performance degradation.
• Supervised learning approaches are limited in real-world scenarios with abundant unlabeled data.

Purpose of the Study:

• To propose a Deep Cross-Subject Adaptation Decoding (DCAD) framework to accurately decipher brain states from fMRI signals.
• To address the challenges of individual differences and data distribution shifts in cross-subject fMRI decoding.
• To enable effective brain state decoding using unlabeled target data.

Main Methods:

• Developed a volume-based 3D feature extraction architecture to learn common spatiotemporal features.
• Employed an unsupervised domain adaptation (UDA) method to minimize the distance between source and target data distributions.
• Evaluated the DCAD framework on the Human Connectome Project (HCP) task-fMRI dataset.

Main Results:

• Achieved state-of-the-art decoding performance with mean accuracies of 81.9% (4 brain states) and 84.9% (9 brain states) on a working memory task.
• Demonstrated that UDA effectively mitigates the impact of data distribution shifts.
• Showcased the framework's ability to improve cross-subject decoding without relying on target domain annotations.

Conclusions:

• The DCAD framework offers a robust solution for decoding cognitive states across subjects using fMRI.
• Unsupervised domain adaptation is a powerful technique for enhancing cross-subject decoding performance in neuroimaging.
• This approach provides a valuable tool for analyzing unlabeled fMRI data and understanding brain function.