Uncovering Brain Dynamics from Spatiotemporal Functional Imaging Via Recursive Latent Constraints and Shared Feature Consistency
Abstract
Purpose
Accurately decoding the spatiotemporal complexities of the human brain requires capturing the intrinsic physical geometry of functional imaging data. Traditional dimensionality reduction methods often struggle with the non-linear, time-varying nature of the BOLD signal, failing to preserve coherence during 4D acquisition. We developed NeuroVis, a deep manifold learning framework for functional imaging that mines spatiotemporal correlations through a recursive outer-inner latent constraint mechanism to enable high-fidelity visualization of neurocognitive trajectories and early detection of aberrant functional states.
Methods
NeuroVis optimizes the representation of 4D functional volumes by integrating spatiotemporal correlation discovery into the latent space. The framework enforces a dual-constraint mechanism: outer refinement to align global temporal dynamics and inner refinement to ensure local spatial feature consistency across voxels. This acts as a topological guide to prevent trajectory fragmentation caused by imaging noise. We validated the framework using the Sherlock naturalistic fMRI dataset (N=16). To evaluate clinical utility, we simulated "progressive cognitive disruption" by injecting linear drift anomalies into real fMRI data, testing the model's sensitivity to pre-symptomatic shifts in functional imaging signals.
Results
NeuroVis demonstrated substantial improvements in trajectory continuity and subject identifiability over baselines (T-PHATE, CEBRA, BCNE), achieving a KNN classification accuracy >0.65. The embeddings revealed a consistent "shared neurocognitive manifold" across subjects. In anomaly detection, NeuroVis identified functional signal deviations prior to overt clinical states, achieving an AUC of 0.964 and a strong correlation (r=0.789) with ground-truth progression. Crucially, the framework provided an "early warning" signal (Lead Time >30 TRs), detecting dynamical regime shifts that traditional sliding-window analyses failed to resolve.
Conclusion
By leveraging recursive spatiotemporal constraints, NeuroVis provides a high-fidelity window into dynamic functional brain states. Its ability to quantify feature consistency and preemptively detect imaging trajectory deviations suggests significant potential for the development of quantitative imaging biomarkers for neurodegenerative disorders.