Image categorization from functional magnetic resonance imaging using functional connectivity.

Area of Science:

• Neuroscience
• Cognitive Neuroscience
• Machine Learning in Neuroscience

Background:

• Previous functional magnetic resonance imaging (fMRI) studies decoded image categories using voxel activity patterns.
• The contribution of whole-brain functional connectivity (FC) patterns to category classification remains largely unexplored.

Purpose of the Study:

• To investigate whole-brain FC patterns for classifying image categories (cats, faces, houses, vehicles).
• To evaluate the performance of machine learning models (SVM, Random Forest) in decoding these FC patterns.
• To examine the impact of window length on FC pattern robustness for neural decoding.

Main Methods:

• fMRI data from healthy adults viewing 4 image categories.
• Whole-brain FC patterns were extracted using varying window lengths (24s, 48s).
• Machine learning classifiers (Support Vector Machine, Random Forest) were employed for classification.

Main Results:

• Classification accuracies of 74% (within-subject) and 80% (between-subject) were achieved, significantly above chance (50%).
• Random Forest outperformed Support Vector Machine.
• Longer window lengths (48s) yielded better classification results than shorter ones (24s).

Conclusions:

• Whole-brain FC patterns effectively predict image categories with high accuracy.
• The findings highlight novel mechanisms of categorical information representation in large-scale brain networks.
• This approach offers a promising avenue for understanding neural representations of visual stimuli.