Transformer³: A Pure Transformer Framework for fMRI-Based Representations of Human Brain Function.

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Neuroscience

Background:

• Effective representation learning is crucial for neuroimage-based individualized predictions.
• Existing studies on functional MRI (fMRI) often analyze only one or two types of interdependencies (sample-wise, spatial, temporal), limiting information extraction.
• A comprehensive approach is needed to leverage all three interdependencies for robust brain function representation.

Purpose of the Study:

• To develop a novel framework, Transformer³, that fully utilizes sample-wise, spatial, and temporal interdependencies in fMRI data.
• To enhance the accuracy of individualized predictions based on brain function representations.
• To provide a versatile tool for representation learning on multivariate time-series data.

Main Methods:

• Proposed a pure transformer-based framework named Transformer³.
• Incorporated three specialized transformer modules: Batch Transformer (sample-wise), Region Transformer (spatial), and Time Transformer (temporal).
• Leveraged the inherent ability of transformers to capture complex interdependencies within the input data.

Main Results:

• Demonstrated the effectiveness of Transformer³ on age, IQ, and sex prediction tasks using two public fMRI datasets.
• Validated the framework's ability to capture comprehensive interdependencies within fMRI data.
• Showcased the potential for Transformer³ in various multivariate time-series representation learning applications.

Conclusions:

• Transformer³ effectively extracts representations of human brain function by leveraging all three types of interdependencies.
• The framework's hypothesis-free nature allows broad applicability to multivariate time-series data.
• The pure transformer architecture facilitates interpretability of predictive model driving factors.