Whole-Brain Task fMRI Decoding Using Stage-Wise Residual-Optimized 3D ConvNeXt With Layer-Global Response...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Decoding brain states from functional magnetic resonance imaging (fMRI) is crucial for cognitive neuroscience and clinical applications.
• Current deep learning models face challenges in balancing task generalization, spatial detail, and interpretability in fMRI analysis.

Purpose of the Study:

• To introduce a novel 3D ConvNeXt framework for whole-brain task-fMRI decoding.
• To improve the accuracy, efficiency, and interpretability of fMRI decoding models.

Main Methods:

• Developed a 3D ConvNeXt framework incorporating layer-global response normalization (LN-GRN) and stage-wise residual connections.
• Evaluated the model on the Human Connectome Project dataset across seven cognitive domains.
• Utilized feature diversity analyses and uniform manifold approximation and projection (UMAP) for clustering and saliency mapping.

Main Results:

• The proposed framework consistently outperformed conventional and specialized 3D fMRI architectures.
• LN-GRN improved feature separability, and restricted residual connections enhanced efficiency without sacrificing accuracy.
• Saliency mapping revealed neuroanatomically meaningful activation patterns, confirming model interpretability.

Conclusions:

• The 3D ConvNeXt framework offers a robust, efficient, and interpretable solution for fMRI decoding, even with limited data.
• The model provides neuroscientific insights by linking predictions to functional brain anatomy.
• This approach shows significant promise for advancing cognitive neuroscience and clinical neuroimaging.