Wee Chong-Yaw, Yap Pew-Thian, Zhang Daoqiang, Wang Lihong, Shen Dinggang
Image Display, Enhancement, and Analysis (IDEA) Laboratory, Biomedical Research Imaging Center (BRIC) and Department of Radiology, University of North Carolina at Chapel Hill, Chapel Hill, NC, 27599, USA,
Brain Struct Funct. 2014 Mar;219(2):641-56. doi: 10.1007/s00429-013-0524-8. Epub 2013 Mar 7.
Emergence of advanced network analysis techniques utilizing resting-state functional magnetic resonance imaging (R-fMRI) has enabled a more comprehensive understanding of neurological disorders at a whole-brain level. However, inferring brain connectivity from R-fMRI is a challenging task, particularly when the ultimate goal is to achieve good control-patient classification performance, owing to perplexing noise effects, curse of dimensionality, and inter-subject variability. Incorporating sparsity into connectivity modeling may be a possible solution to partially remedy this problem since most biological networks are intrinsically sparse. Nevertheless, sparsity constraint, when applied at an individual level, will inevitably cause inter-subject variability and hence degrade classification performance. To this end, we formulate the R-fMRI time series of each region of interest (ROI) as a linear representation of time series of other ROIs to infer sparse connectivity networks that are topologically identical across individuals. This formulation allows simultaneous selection of a common set of ROIs across subjects so that their linear combination is best in estimating the time series of the considered ROI. Specifically, l 1-norm is imposed on each subject to filter out spurious or insignificant connections to produce sparse networks. A group-constraint is hence imposed via multi-task learning using a l 2-norm to encourage consistent non-zero connections across subjects. This group-constraint is crucial since the network topology is identical for all subjects while still preserving individual information via different connectivity values. We validated the proposed modeling in mild cognitive impairment identification and promising results achieved demonstrate its superiority in disease characterization, particularly greater sensitivity to early stage brain pathologies. The inferred group-constrained sparse network is found to be biologically plausible and is highly associated with the disease-associated anatomical anomalies. Furthermore, our proposed approach achieved similar classification performance when finer atlas was used to parcellate the brain space.
利用静息态功能磁共振成像（R-fMRI）的先进网络分析技术的出现，使得在全脑水平上对神经疾病有更全面的理解成为可能。然而，从R-fMRI推断脑连接性是一项具有挑战性的任务，特别是当最终目标是实现良好的对照-患者分类性能时，这是由于令人困惑的噪声效应、维度诅咒和个体间差异所致。将稀疏性纳入连接性建模可能是部分解决此问题的一种可能方法，因为大多数生物网络本质上是稀疏的。然而，稀疏性约束在个体层面应用时，将不可避免地导致个体间差异，从而降低分类性能。为此，我们将每个感兴趣区域（ROI）的R-fMRI时间序列表述为其他ROI时间序列的线性表示，以推断个体间拓扑相同的稀疏连接网络。这种表述允许同时选择跨受试者的一组共同ROI，以便它们的线性组合在估计所考虑ROI的时间序列方面最佳。具体而言，对每个受试者施加l1范数以滤除虚假或不重要的连接，从而产生稀疏网络。因此，通过使用l2范数的多任务学习施加组约束，以鼓励跨受试者一致的非零连接。这种组约束至关重要，因为所有受试者的网络拓扑相同，同时仍通过不同的连接值保留个体信息。我们在轻度认知障碍识别中验证了所提出的建模方法，取得的有前景的结果证明了其在疾病特征描述方面的优越性，特别是对早期脑病变具有更高的敏感性。推断出的组约束稀疏网络在生物学上是合理的，并且与疾病相关的解剖异常高度相关。此外，当使用更精细的图谱对脑空间进行划分时，我们提出的方法实现了类似的分类性能。
