Structurally-constrained encoding framework using a multi-voxel reduced-rank latent model for human natural vision.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Biomedical Engineering

Background:

• Voxel-wise visual encoding models using Convolutional Neural Networks (CNNs) predict human brain activity from fMRI signals.
• Existing CNN models mimic visual cortex hierarchy but lack brain-inspired accuracy for biomedical data prediction.
• There is a need for improved models that capture inter-voxel relationships for more precise brain response prediction.

Purpose of the Study:

• To propose a novel brain-inspired model, the Structurally Constrained Multi-Output (SCMO) module, for enhanced prediction of brain activity.
• To incorporate homologous correlations between voxels within cortical regions to improve response prediction accuracy.
• To assess the predictive performance of the SCMO module using different feature extraction techniques.

Main Methods:

• Developed the SCMO module to leverage population activity and collective voxel behavior for predicting individual voxel-wise BOLD responses.
• Created a structure matrix within the SCMO module to represent voxel-to-voxel interactions within visual regions.
• Evaluated the SCMO module's predictive performance using two feature extraction methods: a recurrent CNN and the AlexNet model.

Main Results:

• The SCMO module demonstrated reliable predictive ability for brain responses across multiple visual areas.
• The proposed framework outperformed benchmark models in terms of prediction stability and feature coherency.
• Analysis revealed the module's capacity to capture underlying voxel-to-voxel interactions within cortical regions.

Conclusions:

• The SCMO module offers a significant advancement in predicting brain activity from fMRI data.
• This brain-inspired approach enhances prediction accuracy by modeling regional voxel correlations.
• The SCMO module provides a more stable and coherent method for visual encoding compared to existing models.