Summary of reviews
- Emma Robinson: 2 (2) View Review 1
- Anonymous reviewer: -1 (3) View Review 2
- Marco Lorenzi: 1 (2) View Review 3
Review 1 (Emma Robinson)
- Reviewer's confidence: 2 (Knowledgable)
- Overall recommendation: 2 (Accept)
SUMMARY
This is an interesting paper that looks at different spectral clustering techniques for identifying functional network communities.
STRENGTHS
In general the paper is very clearly motivated and well written. The methods are solid. The results are quite thorough.
SHORTCOMINGS
I wonder why the authors use an anatomical parcellation to generate their functional networks. In many regions such of the brain (such as the frontal lobe) the relationship between cortical folding and function is not clear, this is why there is so much effort to generate functionally driven segmentations. This may have an impact on why some schemes struggle to differentiate the orbito-frontal from the default mode. It is also true that the cortical folding patterns are highly variable across subjects in higher-order brain regions so it might be a registration issue.
In terms of joint spectral clustering approaches I wonder whether the authors have thought of adopting the joint decomposition approach used by Lombaert et al [1] to clustering. This has been considered for functional segmentation schemes here [2]
Some of the results are quite qualitative. I don’t agree that there are big gaps in the eigenspectrum.
[1] Lombaert, Herve, Jon Sporring, and Kaleem Siddiqi. "Diffeomorphic spectral matching of cortical surfaces." International Conference on Information Processing in Medical Imaging. Springer Berlin Heidelberg, 2013.
[2] Arslan, Salim, Sarah Parisot, and Daniel Rueckert. "Joint spectral decomposition for the parcellation of the human cerebral cortex using resting-state fMRI." International Conference on Information Processing in Medical Imaging. Springer International Publishing, 2015.
CONSTRUCTIVE FEEDBACK
In terms of the results I think it would be better to define what is meant by “nicely separated” or at least make this more clear from the figures which are difficult to interpret. Some more information on how is the Dice overlap calculated would be nice for example is it calculated from overlap of regions defined in figures 2 and 3? What is the interpretation of the Dice overlap results, why is it not consistent between 5 and 8 clusters? What is the neuroscientific interpretations or implications for the use of these techniques for better understanding how the brain works.
I have a few points on clarity: in the preliminaries I don’t think it is helpful to refer to the Rayleigh coefficient unless you define what it is, and why it is relevant. It is probably sufficient to just refer to the Shi and Malik paper and quote the generalized solution as there is not enough space to expand on the full derivation.
Review 2 (Anonymous)
- Reviewer's confidence: 3 (Expert)
- Overall recommendation: -1 (Probably reject)
SUMMARY
The paper evaluates a set of clustering methods on resting-state fMRI data. The authors assess the stability of cluster labelings of the brain resulting from different clustering approaches.
STRENGTHS
The empirical evaluation makes sense, and assesses a relevant metric for clusters.
SHORTCOMINGS
The paper would benefit a lot if the authors would relate the clusters identified by the discussed approaches with some of the many existing brain pracellation results that are based on resting state fMRI data. Those include e.g.,
Yeo, BT Thomas, et al. "The organization of the human cerebral cortex estimated by intrinsic functional connectivity." Journal of neurophysiology 106.3 (2011): 1125-1165.
Wang, Danhong, et al. "Parcellating cortical functional networks in individuals." Nature Neuroscience (2015).
or several papers by Varoquaux et al.
Review 3 (Marco Lorenzi)
- Reviewer's confidence: 2 (Knowledgable)
- Overall recommendation: 1 (Probably Accept)
SUMMARY
This paper investigates the use of multiple-views spectral clustering techniques to identify common functional networks in a group-wise functional connectivity data.
The approach is inspired by [14], in which conventional clustering approaches are applied to the convex combination of the weights matrices of Laplacian associated to different graphs, in order to define group-wise functional clusters.
The authors evaluate different strategies for combining the weight matrices (MVSC, MVSCW, and AASC), and compare the proposed methods to JDL, a different clustering techniques still based on the discrete representation of functional connectivity maps.
Experimental results are carried to investigate the proposed techniques in terms of 1) providing meaningful functional clustering in group-wise data, 2) stability of the obtained clusters, and 3) computational time.
The results suggest that the proposed techniques are able to highlight some meaningful functional patterns corresponding to known brain functional networks, although the interpretability of the results is generally not straightforward, especially when the number of clusters increases. I found the stability analysis of the functional clustering quite insightful. The results obtained for 5 clusters seem to be quite stable with ~100 observations, while in the case of 8 clusters the stability drops considerably with all the techniques.
Given that I’m not an expert in discrete methods for functional connectivity analysis, I think that this study proposes an interesting application of spectral clustering techniques to the problem of group-wise brain functional analysis, and it would be a valuable contribution to the workshop. As a general comment, I would appreciate a comparison with respect to classical analysis methods such as ICA in order to appreciate the real added value of this kind of approaches.
STRENGTHS
- Interesting application of spectral clustering approaches
- Testing of different weighting strategies and comparison with other approaches propose in the literature
- Stability analysis via split-half cross-validation
SHORTCOMINGS
- Notation is not always clear and well defined (e.g. Section 2).
- It is quite concerning that (AASC) generally leads to the worst results. There might be a problem in the weights optimisation, although this part is not clearly explained in the paper.
- The overall method is based on a number of heuristics and approximations, such as the suppression of negative correlations weights, the choice of number clusters, cluster matching in cross-validation based on dice score similarity… I wonder to what extent this can globally affect stability and robustness of the results.
- I think that the interpretation of the results still remains an issue, and is highly dependent on the parameter choice. The methods seem to be not very reliable in detecting a variety of functional networks
CONSTRUCTIVE FEEDBACK
I believe that the proposed approaches should be compared to standard methods for brain functional analysis, such as the ones based on ICA. This might shed more lights on the interpretability and reliability of the results.
Similarly, I would recommend to test the proposed approaches on synthetic data.