Diverse Task Classification from Activation Patterns of Functional Neuro-Images Using Feature Fusion Module

Area of Science:

• Neuroimaging
• Machine Learning
• Cognitive Neuroscience

Background:

• Decoding brain activity using functional magnetic resonance imaging (fMRI) is an emerging field.
• Current methods struggle with multi-task classification due to limited datasets and unaddressed challenges.
• Existing studies predominantly focus on single neuronal activity tasks.

Purpose of the Study:

• To address limitations in multi-task classification within fMRI.
• To introduce a novel dataset encompassing diverse cognitive, behavioral, and affective tasks.
• To propose an automatic multi-task classification (MTC) strategy with a feature fusion module (FFM).

Main Methods:

• Collected and released a new fMRI dataset with cognitive, behavioral, affective tasks, and resting state data.
• Developed a feature fusion module (FFM) to combine deep features and time-frequency representations for task signature creation.
• Evaluated MTC performance using diverse deep learning models and analyzed their complementarity.

Main Results:

• The FFM created a superior feature space for representing task characteristics compared to individual feature types.
• The proposed MTC strategy achieved higher classification accuracy than existing benchmarks.
• The new dataset and code are publicly available for research.

Conclusions:

• The study presents a novel dataset and an effective FFM-based MTC strategy for fMRI.
• The findings demonstrate improved performance in decoding multiple brain activity tasks.
• This work facilitates further advancements in fMRI decoding and multi-task learning.