New 3D ConvNeXt Framework Boosts Whole-Brain fMRI Decoding with Enhanced Interpretability

The ability to accurately decode brain states from functional magnetic resonance imaging (fMRI) is crucial for understanding the human brain and developing new medical tools. However, current deep learning methods often struggle to balance accuracy, generalizability, and interpretability, particularly in large-scale or clinical studies. This limits their real-world impact and the depth of neuroscientific insights they can provide.

To overcome these challenges, researchers have introduced a new 3D ConvNeXt framework specifically designed for whole-brain task fMRI decoding. This innovative model incorporates layer-global response normalization (LN-GRN) to optimize feature scaling and utilizes stage-wise residual connections, enhancing computational efficiency without sacrificing accuracy. When tested on the extensive Human Connectome Project dataset, the framework consistently outperformed both conventional neural networks and specialized 3D MRI architectures across various cognitive tasks.

Key to its success, the LN-GRN component significantly improved the distinctiveness of features, while strategically limiting residual connections to earlier stages maintained accuracy with reduced complexity. Advanced analyses, including uniform manifold approximation and projection-based clustering, confirmed superior class separation. Furthermore, saliency mapping revealed activation patterns that aligned with known brain organization, providing neuroanatomically meaningful insights.

As lead author Lim notes in the paper, "These findings demonstrate that our proposed framework provides robust, efficient, and interpretable fMRI decoding, even under conditions of limited data." Beyond its methodological innovations, the model provides valuable neuroscientific insights by directly linking its predictions to functional brain anatomy. This powerful new approach holds strong promise for advancing cognitive neuroscience research and has significant potential for clinical neuroimaging applications, including the early diagnosis and characterization of neurological disorders.