Task sub-type states decoding via group deep bidirectional recurrent neural network.

Area of Science:

• Neuroimaging
• Machine Learning
• Cognitive Neuroscience

Background:

• Functional magnetic resonance imaging (fMRI) is crucial for decoding brain states during cognitive tasks.
• Existing methods often fail to fully exploit temporal dependencies in fMRI data due to limited machine learning capacities.
• Current training strategies can also hinder the effective capture of brain activity contrasts between tasks.

Purpose of the Study:

• To propose a novel method for fine-grain brain state decoding from fMRI data.
• To address limitations in temporal dependency modeling and training sample organization in existing brain decoding techniques.
• To improve the accuracy and interpretability of brain state decoding.

Main Methods:

• Developed a Group Deep Bidirectional Recurrent Neural Network (Group-DBRNN) model.
• Introduced a training sample organization strategy with group-task sample generation and a multiple-scale random fragment strategy (MRFS).
• Incorporated bidirectional stacked RNNs and a multi-task interaction layer (MTIL) to capture temporal dependencies and task-relevant contrasts.

Main Results:

• Achieved an average decoding accuracy of 94.7% across 23 fine-grain sub-type states on the Human Connectome Project fMRI dataset.
• Demonstrated effective capture of temporal dependency and task-relevant brain activity contrast.
• Intermediate features learned by the model showed strong discriminability and inter-subject alignment.

Conclusions:

• The Group-DBRNN model significantly advances fine-grain brain state decoding from fMRI data.
• The proposed methods for sample organization and model architecture effectively address limitations in existing approaches.
• The model's ability to capture temporal dynamics and task contrasts offers valuable insights into brain function.