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

Decoding brain states from task-based functional magnetic resonance imaging (fMRI) is critical for advancing cognitive neuroscience and developing reliable clinical applications. Existing deep learning methods, however, often struggle to balance task generalization, spatial fidelity, and neuroscientific interpretability, limiting their effectiveness in large-scale and clinical studies. To address these challenges, we introduce a 3D ConvNeXt framework designed explicitly for whole-brain task fMRI decoding. The model integrates layer-global response normalization (LN-GRN) for improved feature scaling and employs stage-wise residual connections to enhance computational efficiency without compromising accuracy. Evaluated on the human connectome project dataset covering seven cognitive domains, the proposed framework consistently outperformed conventional convolutional neural networks, and specialized 3D magnetic resonance imaging architectures across all tasks. LN-GRN enhanced feature separability, while restricting residual connections to Stages 1-3 preserved accuracy with reduced complexity. Feature diversity analyses and uniform manifold approximation and projection-based clustering confirmed superior class separation, and saliency mapping revealed neuroanatomically meaningful activation patterns aligned with known brain organization. These findings demonstrate that our proposed framework provides robust, efficient, and interpretable fMRI decoding, even under conditions of limited data. Beyond methodological contributions, such as optimal residual connection placement and LN-GRN integration, the model provides neuroscientific insights by linking predictions to functional brain anatomy. This approach holds strong promise for advancing cognitive neuroscience research and supporting clinical neuroimaging applications, including early diagnosis and characterization of neurological disorders.