Analyzing Neuroimaging Data Through Recurrent Deep Learning Models.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Deep learning (DL) models face challenges with high-dimensional neuroimaging data, including low sample sizes and complex spatiotemporal dependencies.
• Existing DL models often function as "black boxes," limiting insights into the relationship between cognitive states and brain activity.

Purpose of the Study:

• Introduce the DeepLight framework for analyzing whole-brain functional Magnetic Resonance Imaging (fMRI) data using long short-term memory (LSTM) models.
• Enhance the interpretability of DL models in neuroimaging by adapting layer-wise relevance propagation (LRP) for voxel-wise contribution analysis.
• Enable the study of cognitive state-brain activity associations across multiple data granularities, from group-level to single time points.

Main Methods:

• DeepLight processes fMRI data by segmenting volumes into sequential axial slices for LSTM analysis.
• The framework integrates Layer-wise Relevance Propagation (LRP) to decompose decoding decisions into voxel-specific contributions.
• The approach was validated on a large Human Connectome Project fMRI dataset.

Main Results:

• DeepLight demonstrated superior performance compared to conventional univariate and multivariate fMRI analysis methods in decoding cognitive states.
• The framework successfully identified physiologically relevant brain regions associated with specific cognitive states.
• DeepLight effectively captured fine-grained spatiotemporal variability in brain activity across individual fMRI samples.

Conclusions:

• DeepLight offers a powerful and interpretable DL approach for whole-brain fMRI analysis.
• The framework advances the understanding of cognitive state-brain activity relationships by providing detailed insights into neural mechanisms.
• DeepLight shows significant potential for future neuroimaging research and clinical applications.