Abstract
Analysis and interpretation of neuroimaging data often require one to divide the brain into a number of regions, or parcels, with homogeneous characteristics, be these regions defined in the brain volume or on the cortical surface. While predefined brain atlases do not adapt to the signal in the individual subject images, parcellation approaches use brain activity (e.g., found in some functional contrasts of interest) and clustering techniques to define regions with some degree of signal homogeneity. In this work, we address the question of which clustering technique is appropriate and how to optimize the corresponding model. We use two principled criteria: goodness of fit (accuracy), and reproducibility of the parcellation across bootstrap samples. We study these criteria on both simulated and two task-based functional Magnetic Resonance Imaging datasets for the Ward, spectral and k-means clustering algorithms. We show that in general Ward’s clustering performs better than alternative methods with regard to reproducibility and accuracy and that the two criteria diverge regarding the preferred models (reproducibility leading to more conservative solutions), thus deferring the practical decision to a higher level alternative, namely the choice of a trade-off between accuracy and stability.
Highlights
Brain parcellations divide the brain’s spatial domain into a set of non-overlapping regions or modules that show some homogeneity with respect to information provided by one or several image modalities, such as cyto-architecture, anatomical connectivity, functional connectivity, or task-related activation
SIMULATIONS Figure 2 displays the selected K value, based on data smoothed with a kernel of size 0.5 voxel, and under spatial jitter of 1 voxel isotropic; given that Ktrue = 5 it shows that BIC tends to select too large number of clusters, while, on the opposite, reproducibility, measured via bootstrapped Adjusted mutual information (AMI), is conservative; cross-validated log-likelihood shows an intermediate behavior, as it is conservative for spectral clustering and anti-conservative for k-means
The right model is not recovered in general, because the true clustering is not in the solution path of the different methods, or because model selection fails to recover the right number of parcels
Summary
Brain parcellations divide the brain’s spatial domain into a set of non-overlapping regions or modules that show some homogeneity with respect to information provided by one or several image modalities, such as cyto-architecture, anatomical connectivity, functional connectivity, or task-related activation. Brain parcellations are often derived from specific clustering algorithms applied to brain images. Such approaches are generally useful because the voxel sampling grid of the reference space, e.g., the MNI template, is most often at a higher resolution than the brain structures of interest, or at a scale that is too fine for the problem under investigation, yielding an excessive number of brain locations and correlated data. The obvious limitation of ROI-based analysis is that the signal present outside the region under consideration is ignored a priori; as a consequence, the results depend heavily on the choice of this ROI, which may not fit well the new data. In the hypothesis testing framework, the smaller number of tests performed may, increase the power of the analysis
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have