Abstract
Dynamic functional connectivity analysis provides valuable information for understanding brain functional activity underlying different cognitive processes. Traditionally, methods like sliding-window analysis or statistical change-point detection have been used to segment fMRI scans into functional states. However, these often rely on rigid assumptions, such as Gaussian distributions, which may not suit complex fMRI data. This study introduces a data-driven framework using Long Short-Term Memory (LSTM) recurrent neural networks to adaptively detect functional state transitions. By framing change-point detection as an anomaly detection task, the model identifies state shifts based on spikes in prediction error. Experimental results on Human Connectome Project (HCP) data demonstrate that this method effectively identifies transitions in both task-based and resting-state fMRI, outperforming traditional Bayesian approaches.
Introduction
The human brain's functional connectivity (FC) is highly dynamic, shifting as different regions interact to support various cognitive tasks. Mapping these transitions is essential for understanding brain function. While sliding-window methods are common, they are limited by fixed window sizes. Existing change-point detection methods often struggle with the high noise and non-Gaussian nature of fMRI signals. This research proposes a flexible, deep learning-based approach to identify these boundaries without pre-defining signal models.
Methods and Framework
The proposed framework leverages the sequence-modeling strengths of LSTM-RNNs. The process begins with feature extraction using "collaborative sparse brain decomposition" to reduce 4D fMRI data into time courses for 90 functional networks.
An LSTM network with two hidden layers (256 nodes each) is trained to predict the functional profile of the next time point based on the preceding sequence. The core hypothesis is that within a stable functional state, temporal profiles are predictable; however, a transition to a new state will cause a significant spike in prediction error. To manage the low signal-to-noise ratio of fMRI data, the error signal is smoothed using a 1D Gaussian convolutional filter. Change points are then identified as local maxima in this smoothed error signal that exceed a tuned threshold.
Experimental Data
The study utilized data from 490 subjects in the Human Connectome Project (HCP):
- Task fMRI: Motor tasks (e.g., moving hands/feet) and Working Memory tasks (0-back/2-back) provided a "ground truth" for state transitions.
- Resting-state fMRI: Used to evaluate the model's ability to detect spontaneous, unprompted state shifts.
Key Results
The LSTM-based method demonstrated superior performance compared to state-of-the-art Bayesian inference methods. In task-based fMRI, the model accurately matched the timing of task blocks, even when accounting for hemodynamic lag.
Key findings include:
- Higher Accuracy: The LSTM method showed significantly lower distance errors between predicted and actual change points and fewer false positives than the Bayesian baseline.
- Resting-State Validity: In resting-state data, two-sample covariance matrix testing confirmed that the segments identified by the model possessed statistically distinct functional connectivity patterns ($p < 0.05$).
- Robustness: The use of smoothing and thresholding allowed the model to remain effective despite the inherent noise in fMRI signals.
Discussion and Future Work
While the framework is largely data-driven, it requires optimization of the smoothing kernel ($\sigma$) and threshold ($\lambda$). Future research could focus on automating this parameter selection. Additionally, while this study identifies when transitions occur, further work is needed to characterize the biological significance of the identified resting-state segments. Clustering these segments across subjects may reveal universal functional states common across the human population.
Summary updated