Functional Brain Network Estimation with Time Series Self-scrubbing

Weikai Li, Lishan Qiao, Zhengxia Wang, Dinggang ShenbioRxiv 2017 · DOI

TL;DR

This paper introduces a novel method for estimating functional brain networks (FBNs) by simultaneously cleaning noisy time-series data and building the network model. By incorporating a "data quality indicator" into the estimation process, the method adaptively identifies and removes unreliable data points, leading to more accurate FBNs and improved classification of neurological conditions.

What Problem Does This Paper Address?

Functional Brain Networks (FBNs) are crucial for understanding brain function and identifying biomarkers for neurological disorders. However, the accuracy of FBN estimation is heavily compromised by "noisy" or "dirty" data points within the BOLD (Blood-Oxygen-Level Dependent) signals from fMRI scans. These artifacts can arise from various sources, including head motion, mind-wandering during resting-state scans, and even the preprocessing steps themselves. While traditional methods employ preprocessing pipelines and "scrubbing" techniques to clean the data, these approaches are often imperfect, may remove too much data, and are independent of the FBN estimation task, meaning they don't guarantee the remaining data is optimal for network building. This paper tackles the challenge of obtaining cleaner, more reliable FBNs by integrating data cleaning directly into the FBN estimation process.

Methods

The core innovation of this paper is a "self-scrubbing" FBN estimation framework. Instead of a two-step process (clean data, then estimate FBN), they propose a single, integrated approach. This is achieved by:

  • Introducing a Latent Indicator Variable: A binary variable ($v_t$) is introduced for each time point ($t$) in the fMRI time series. This variable acts as an indicator of data quality: $v_t=1$ for "clean" data and $v_t=0$ for "dirty" data. This indicator is not observed directly but is learned alongside the FBN.
  • Extending the Regularized FBN Framework: The standard regularized FBN estimation framework (which balances data-fitting terms with regularization terms encoding prior knowledge about brain organization) is extended to include this latent indicator. The objective function now penalizes the use of "dirty" time points.
  • Alternating Optimization Algorithm: To solve the complex optimization problem involving both the network weights ($W$) and the indicator variables ($V$), an alternating convex search (ACS) method is employed. This involves iteratively updating $W$ while keeping $V$ fixed, and then updating $V$ while keeping $W$ fixed.
  • Relaxation and Regularization for $v_t$: To make the optimization tractable, the binary constraint ($v_t \in \{0,1\}$) is relaxed to a continuous range ($v_t \in [0,1]$). A negative $l_1$-norm regularization term is added to encourage sparsity in the indicator variable, effectively promoting the removal of time points.
  • Partial Correlation as Data-Fitting Term: The paper adopts partial correlation as the data-fitting term, which is known to be more robust than simple correlation in FBN estimation.

Datasets

The proposed method was evaluated on two publicly available datasets:

  • ADNI (Alzheimer's Disease Neuroimaging Initiative): This dataset included 110 participants (51 MCI patients and 59 Normal Controls).
  • NITRC Dataset: This dataset comprised 46 MCI patients and 45 Normal Controls.

The NITRC dataset had undergone a traditional scrubbing operation prior to the experiments, allowing the authors to test their method on both pre-scrubbed and un-scrubbed data.

Key Insights and Ideas

  • Integrated Data Cleaning and FBN Estimation: The most significant insight is that data cleaning and FBN estimation are not independent tasks. By jointly optimizing them, the method can learn which data points are truly beneficial for building an accurate FBN, leading to a more principled and effective approach than sequential scrubbing.
  • Adaptive Data Scrubbing: The "self-scrubbing" mechanism is adaptive. It doesn't rely on fixed thresholds for head motion but learns which time points are problematic based on their impact on the FBN estimation itself. This is crucial because "dirty" points can arise from various, not always obvious, sources.
  • Improved FBN Accuracy: The method demonstrably produces FBNs that are more sensitive to subtle differences, as evidenced by the improved classification performance. The changes observed in specific brain regions (cerebellum, hippocampus, frontal regions) after self-scrubbing suggest these areas are particularly susceptible to motion artifacts and functional noise.
  • Flexibility: The framework is flexible, allowing for the integration of different data-fitting terms and regularization priors, making it adaptable to various FBN estimation scenarios.

Key Results

  • Enhanced Classification Accuracy: The proposed SR+SS method significantly outperformed baseline methods (PC and SR) in classifying Mild Cognitive Impairment (MCI) patients from Normal Controls (NC) on both the ADNI and NITRC datasets. For example, on the ADNI dataset, the accuracy reached 84.31% with SR+SS, compared to 79.09% for PC and 82.35% for SR. Similar improvements were observed for sensitivity and specificity.
  • Effective Noise Removal: Experiments on simulated data showed that the proposed method iteratively identified and removed noisy time points, gradually refining the FBN estimation.
  • Identification of Sensitive Brain Regions: Analysis of the changes in FBNs after self-scrubbing highlighted cerebellum, hippocampus, and frontal regions as areas where connection weights were most affected, suggesting these regions are more prone to noise.

Limitations and Future Work

  • Initial Iteration: The authors acknowledge that in the first iteration, all time points are initially considered "clean," which could lead to a suboptimal regression result (local minimum) if a large number of "dirty" points are present from the outset.
  • Information Loss: The current "hard indicator" approach discards removed points entirely. The authors suggest that potentially useful information from these points might be lost.
  • Future Directions:
    • Developing a "soft" or probabilistic version of the indicator variable for greater flexibility.
  • Exploring the transferability of the self-scrubbing strategy to other FBN estimation models with different data-fitting and regularization terms.
  • Investigating ways to recover potentially useful information from discarded time points.

Summary updated

Abstract

Abstract Functional brain network (FBN) has been becoming an increasingly important measurement for exploring the cerebral working mechanism and mining informative biomarkers for assisting diagnosis of some neurodegenerative disorders. Despite its potential performance in discovering the valuable patterns hidden in the brains, the estimated FBNs are often heavily influenced by the quality of the observed data (e.g., BOLD signal series). In practice, a preprocessing pipeline is usually employed for improving the data quality prior to the FBN estimation; but, even so, some data points in the time series are still not clean enough, possibly including original artifacts (e.g., micro head motion), non-resting functional disturbing (e.g., mind-wandering), and new “noises” caused by the preprocessing pipeline per se. Therefore, not all data points in the time series can contribute to the subsequent FBN estimation. To address this issue, in this paper, we propose a novel FBN estimation method by introducing a latent variable as an indicator of the data quality, and develop an alternating optimization algorithm for scrubbing the data and estimating FBN simultaneously in a single framework. As a result, we can obtain more accurate FBNs with the self-scrubbing data. To illustrate the effectiveness of the proposed method, we conduct experiments on two publicly available datasets to identify mild cognitive impairment (MCI) patients from normal control (NC) subjects based on the estimated FBNs. Experimental results show that the proposed FBN modelling method can achieve higher classification accuracy, significantly outperforming the baseline methods.
BibTeX
@misc{li2017functional,
  author = {Weikai Li and Lishan Qiao and Zhengxia Wang and Dinggang Shen},
  title = {{Functional Brain Network Estimation with Time Series Self-scrubbing}},
  year = {2017},
  doi = {10.1101/191262},
  howpublished = {bioRxiv},
}
Figure