Comprehensive decoding mental processes from Web repositories of functional brain images.

Area of Science:

• Neuroscience
• Cognitive Science
• Machine Learning
• Data Science

Background:

• Associating brain systems with mental processes necessitates statistical analysis of brain activity across diverse cognitive tasks.
• Traditional analyses face a trade-off between the scope of analysis (domain-specific vs. system-level) and accuracy.
• Large-scale repositories of neuroimaging data offer potential for more comprehensive analyses but present challenges in data heterogeneity and curation.

Purpose of the Study:

• To develop a scalable method for decoding cognitive concepts from functional Magnetic Resonance Imaging (fMRI) data.
• To overcome limitations of traditional meta-analyses by leveraging machine learning on a large, open-access dataset.
• To enable broad reverse inferences, linking observed brain activity to specific mental processes.

Main Methods:

• Utilized the largest available repository of fMRI statistical maps from NeuroVault, an open data repository.
• Integrated NeuroVault data with the Cognitive Atlas, an ontology of cognition, to label brain images with relevant concepts.
• Trained neural networks on tens of thousands of fMRI images to predict cognitive labels, decoding over 50 mental process classes.

Main Results:

• Successfully decoded more than 50 classes of mental processes from unseen fMRI studies using trained machine learning models.
• Demonstrated the feasibility of large-scale, image-based meta-analyses with minimal manual data curation.
• Overcame challenges related to data heterogeneity, imbalance, and noise in the training dataset.

Conclusions:

• Machine learning models trained on large-scale fMRI data can accurately decode cognitive concepts.
• This approach facilitates robust reverse inference, linking brain activity patterns to specific mental processes.
• The study highlights the potential of open data repositories and machine learning for advancing cognitive neuroscience research.