Author: Lei Xing, Zixia Zhou 👨🔬
Affiliation: Department of Radiation Oncology, Stanford University, Department of Radiation Oncology, Stanford University, Stanford 🌍
Purpose: Functional brain imaging techniques, such as functional magnetic resonance imaging (fMRI), generate high-dimensional, dynamic data reflecting complex neural processes. However, extracting robust neurocognitive patterns remains challenging due to temporal complexity, spatial interdependencies, and measurement noise. This study introduces a self-supervised deep manifold learning framework, Brain Cognitive Network Embedding (BCNE), to uncover interpretable cognitive and behavioral trajectories, advancing brain research.
Methods: BCNE captures dynamic patterns by leveraging temporospatial correlations, unlike traditional approaches that directly extract patterns from input data. It reformulates multivariate time-series signals into temporospatial grids, preserving temporal continuity and spatial dependencies. A manifold learning approach minimizes the Kullback–Leibler (KL) divergence between high- and low-dimensional distance matrices as a self-supervised loss, optimizing network training. Recursive optimization progressively refines the latent manifold, integrating deeper constraints to uncover nuanced temporal dynamics. BCNE's linear computational complexity ensures scalability, and its self-supervised design enables unbiased pattern discovery without labeled data.
Results: Validated on multiple fMRI BOLD datasets, BCNE delineates scene transitions during movie watching, reveals distinct memory and narrative processing patterns, and traces diverging brain trajectories through skill-learning phases. Comprehensive evaluation includes scene classification, assessing rapid transitions, long-range dependencies, and activity pattern consistency, complemented by a behavioral difference metric to quantify the embeddings’ ability to capture meaningful cognitive transitions. BCNE achieves the highest accuracy, surpassing T-PHATE (SoTA) by 81.74% in scene classification and improving over sevenfold in reflecting behavioral differences (averaged across subjects and ROIs). Compared to UMAP, T-PHATE, and CEBRA, BCNE consistently delivers more coherent embeddings, enhanced stability, and deterministic brain dynamic patterns.
Conclusion: BCNE offers a transformative platform for decoding complex brain signals. By modeling temporospatial dependencies and leveraging robust manifold optimization, it provides groundbreaking insights into neurocognitive processes. Its scalability, deterministic outputs, and self-supervised learning paradigm make it a pivotal tool for neuroscience research and clinical applications.