Abstract
Brain functional network (BFN) analysis is becoming a crucial way to explore the inherent organized pattern of the brain and reveal potential biomarkers for diagnosing neurological or psychological disorders. In so doing, a well-estimated BFN is of great concern. In practice, however, noises or artifacts involved in the observed data (i.e., fMRI time series) generally lead to a poor estimation of BFN, and thus a complex preprocessing pipeline is often used to improve the quality of the data prior to BFN estimation. One of the popular preprocessing steps is data-scrubbing that aims at removing "bad" volumes from the fMRI time series according to the amplitude of the head motion. Despite its helpfulness in general, this traditional scrubbing scheme cannot guarantee that the removed volumes are necessarily unhelpful, since such a step is fully independent to the subsequent BFN estimation task. Moreover, the removal of volumes would reduce the statistical power, and different numbers of volumes are generally scrubbed for different subjects, resulting in an inconsistency or bias in the estimated BFNs. To address these issues, we develop a new learning framework that conducts BFN estimation and data-scrubbing simultaneously by an alternating optimization algorithm. The newly developed algorithm adaptively weights volumes (instead of removing them directly) for the task of BFN estimation. Finally, we validate the proposed method by identifying subjects with mild cognitive impairment (MCI) from normal controls based on the estimated BFNs, achieving an 80.22% classification accuracy, which significantly improves the baseline methods.
Introduction
Functional Magnetic Resonance Imaging (fMRI) has revolutionized our ability to map the functional connectivity of the human brain. By analyzing Brain Functional Networks (BFNs), researchers can identify patterns of interaction between distinct brain regions that serve as biomarkers for various neurological conditions. However, fMRI data is notoriously susceptible to noise, particularly from head motion, which can severely distort BFN estimation.
Traditionally, researchers have relied on "data-scrubbing" to mitigate these artifacts. This process involves identifying and deleting "bad" volumes—time points where motion exceeds a specific threshold—before any network analysis begins. While effective at reducing noise, this "hard scrubbing" approach suffers from significant drawbacks: it reduces statistical power by discarding data, introduces bias because different subjects lose different amounts of data, and operates independently of the final network estimation goal.
The Proposed Method: Adaptive "Soft" Scrubbing
To overcome the limitations of traditional scrubbing, this paper introduces a joint learning framework that performs BFN estimation and data cleaning simultaneously. Instead of a binary decision to keep or discard a volume, the method assigns a continuous weight ($w_t$) to every time point in the fMRI series.
As shown in Figure 1, the framework utilizes an Alternating Optimization (AO) algorithm. The process iterates between two main steps:
- Network Estimation: With weights fixed, the algorithm estimates the network connections using Sparse Representation (SR) to find partial correlations between regions.
- Weight Updating: With the network model fixed, the algorithm updates the weights. A volume's weight is determined by its "fitting error"; if a volume does not align well with the emerging network structure (indicating it is likely noisy), its weight is automatically reduced.
This "soft scrubbing" ensures that every volume contributes to the final model according to its quality, maintaining a consistent data structure across all subjects while minimizing the impact of artifacts.
Experimental Setup and Validation
The proposed method was validated using a dataset from the Neuroimaging Informatics Tools and Resources Clearinghouse (NITRC), consisting of 91 subjects (45 with Mild Cognitive Impairment (MCI) and 46 Normal Controls). The brain was parcellated into 90 Regions of Interest (ROIs) using the Automated Anatomical Labeling (AAL) atlas.
The identification pipeline, illustrated in Figure 3, involves extracting BOLD signals, estimating BFNs via the proposed adaptive weighting, and then using a Leave-One-Out (LOO) cross-validation scheme with a Support Vector Machine (SVM) to classify subjects.
Key Results and Discussion
The proposed method (SR+W) achieved a classification accuracy of 80.22%, significantly outperforming several baseline approaches:
- Pearson’s Correlation (PC): ~70% accuracy.
- Standard Sparse Representation (SR): ~73% accuracy.
- SR with Hard Scrubbing (SR+SS): ~78% accuracy.
Beyond improved accuracy, the model identified key discriminative regions in the temporal gyrus, parahippocampus, and precuneus. These findings align with established clinical knowledge regarding the early stages of Alzheimer’s disease and MCI, suggesting that the adaptively-weighted BFNs capture more biologically relevant information than traditional methods.
Conclusion
By integrating data-scrubbing directly into the BFN estimation process, this framework provides a more robust and flexible way to handle fMRI noise. While the model is sensitive to regularization parameters and currently operates in an unsupervised manner regarding diagnostic labels, it represents a significant step toward more reliable automated diagnosis of neurological disorders. Future work may explore supervised BFN learning to further enhance the discriminative power of these functional maps.
Summary updated