Parcellating the parcellation issue - a proof of concept for reproducible analyses using Neurolibre

Pierre bellec 🤔 💻 👀

Saâd Jbabdi 👀

R. Cameron Craddock 👀

1Introduction¶

Back in 2017, we were invited as guest editors for the journal “neuroimage”, and organized a special issue on the topic of brain parcellation and segmentation Craddock et al., 2018. After months of receiving submissions and going through the stages of peer review for all articles, there were 38 papers accepted for publication. We were left with the difficult task of writing a short editorial summarizing concisely this fairly large snapshot of a vibrant area of neuroimaging research. Quickly we realized that we needed to group the papers into more specific categories than the general umbrella of parcellation and segmentation.

Parcellation, segmentation (or clustering and community detection) are just different names for generic techniques aimed at finding broad categories in the data, while abstracting away less relevant finer-grained information. So when facing the problem of drawing categories of papers for the editorial, it was only natural to apply a parcellation analysis. In other words, we decided to parcellate the parcellation (special) issue. This jupyter book will allow interested readers to reproduce the short analysis that went into building the categories. You can read the final editorial here (sadly behind a paywall).

This reproducible analysis is possible thanks to the following projects:

• The python programming language.

• Jupyter book is used to generate the html pages, and is itself based on sphynx.

• The NeuroLibre is hosting this jupyter book.

• The scipy scientific library.

• The seaborn visualization library.

• The wordcloud library.

2Text mining¶

2.1List of papers¶

We first assembled the title, the name of the corresponding author, and the abstract for all the articles into a tabular-separated values (tsv) file, which we publicly archived on Figshare. We use the Repo2Data tool developed by the NeuroLibre team to collect these data and include them in our reproducible computational environment.

2.2Word features¶

For each paper, we used scikit-learn Kramer, 2016 to extract a bag of words representation for each abstract, picking on the 300 most important terms seen across all articles based on a term frequency-inverse document frequency (tf-idf) index. Following that, a special value decomposition was used to further reduce the dimensionality of the abstracts to 10 components. We ended up with a component matrix of dimension 38 (articles) times 10 (abstract text components). The distribution of each of the 38 articles across the 10 components is represented below. Note how some articles have particular high loadings on specific components, suggesting these may capture particular topics. Rather than visually inspect the component loadings to group paper ourselves, we are going to resort to an automated parcellation (clustering) technique.

2.3Parcellate the papers¶

Now that the content of each paper has been condensed into only 10 (hopefully informative) numbers, we can run these features into a trusted, classic parcellation algorithm: Ward’s agglomerative hierarchical clustering, as implemented in the scipy library. We cut the hierarchy to extract 7 “paper parcels”, and also use the hierarchy to re-order the papers, such that similar papers are close in order, as illustrated in a dendrogram representation.

2.4Similarity matrix¶

So, to get a better feel of the similarity between papers that was fed into the clustering procedure, we extracted the 38x38 (papers x papers) correlation matrix across features. Papers are re-ordered in the matrix according to the above hierarchy. Each “paper parcel” has been indicated by a white square along the diagonal, which represents the similarity measures between papers falling into the same parcel.

2.5Word cloud¶

Now, each paper of the special issue has been assigned to one and only one out of 7 possible “paper parcel”. For each paper parcel, we can evaluate which words contribute more to the dominant component associated with that parcel.

2.6Categories¶

Thanks to the word clouds, these simple data-driven categories turned out to be fairly easily interpretable. For example, the word cloud of the category number 4 features prominently words like “white”, “matter” and “bundles”. If we examine the exact list of papers included in this category, we see that it is composed of four papers, which all considered parcels derived from white matter bundles with diffusion imaging.

We can also check the distribution of component loadings for this category alone. As expected, there is a certain similarity in the component loadings for these papers, in particular along component 4:

3Conclusion¶

By checking the word clouds and the titles, the categories of papers are pretty clear. The editorial we wrote was based on this analysis, and we went through each paper category in a separate section. As a side note, although this analysis does not really have scientific value, apart from sparing us from manually assigning papers to categories, it is concise and features a range of visualizations. So we chose to use the analysis as a proof of concept for reproducible publications using the Neurolibre preprint service, which may be the medium where you are reading these lines.