An interactive meta-analysis of MRI biomarkers of myelin


This Jupyter Notebook explores an important aspect of quantitative magnetic resonance imaging (qMRI): validation. Focusing specifically on myelin measures, we show the results of our meta-analysis comparing quantitative MRI with histology. The full preprint is available on bioRxiv.

Why myelin? Myelin is a key component of the central nervous system. The myelin sheaths insulate axons with a triple effect: allowing fast electrical conduction, protecting the axon, and providing trophic support. The conduction velocity regulation has become an important research topic, with evidence of activity-dependent myelination as an additional mechanism of plasticity. Myelin is also relevant from a clinical perspective, given that demyelination is often observed in several neurological diseases such as multiple sclerosis.

How are qMRI measures validated? Similarly to other qMRI biomarkers, MRI-based myelin measurements are noisy, indirect, and might be affected by other microstructural features. Assessing the accuracy of such measurements, as well as their sensitivity to change, is essential for their translation into clinical practice. That is why histological validation is necessary. The most common validation approach is based on acquiring MR data from in vivo or ex vivo tissue and then comparing those data with the related samples analysed using histological techniques.

Why a meta-analysis? So far, a long list of studies have looked at MRI-histology comparisons, each of them focusing on a specific pathology and a few MRI measures. Despite these numerous studies, there is still an ongoing debate on what MRI measure should be used to quantify myelin and as a consequence there is a constant methodological effort to propose new measures. We believe that this debate would benefit from a quantitative analysis of all the findings published so far, specifically addressing inter-study variations and prospects for future studies, something that is currently missing from the literature.


A bit more about this Jupyter Notebook

The main idea for this Jupyter Notebook is to let you interactively explore our dataset. For this reason, in the next pages you will find brief descriptions of what has been done, but the figures (realized with plotly) are the main content. You will not find sections discussing or interpreting these results: for that, please check the preprint.


Selecting the studies

First, how were the studies selected? We used the Medline database and retrieved all the records mentioning (1) myelin, (2) MRI and (3) histology (or a related technique). The full list of keywords is provided in the preprint. The following Sankey diagram shows the screening process: you can hover with the mouse on each block and connection to see details about the number of studies and exclusion criteria.

We identified 58 studies reporting quantitative comparisons between MRI and histology: these included a variety of methodological choices and experimental conditions, in terms of tissue type (brain, spinal cord, peripheral nerve), condition (in vivo, ex vivo, in situ), species (human, animal), pathology model, and many more. A glimpse of these subdivisions is provided in the following treemap. You can click on each box to expand the related category, and for each study you can find out more details and the link to the original paper.

A closer look

Given the number of different variables influencing the results, we decided to focus only on brain studies. As we needed to take into account the sample size for quantitative comparisons, we also further selected only the studies that reported both the number of subjects and the number of ROIs (regions of interest) considered for correlation purposes. This further screening led us to 43 studies. For these studies we wanted to quantitatively evaluate the reported effect size taking into account the respective samples sizes: we chose the coefficient of determination R2, as it was the most common quantitative result we could obtain from these studies. To have a look at both sample size and effect size for each measure, we prepared an interactive bubble chart, where the size of each bubble is proportional to the sample size. You can hover on the bubbles to obtain additional details.

To provide a different way to explore sample size and effect size, we also prepared another treemap, where the studies are organised by measures. For each study, the area of its box is proportional to the sample size, while the color represents the related coefficient of determination.

Using meta-analysis tools

Can we express quantitatively what we observed in the previous plots? This is where the meta-analysis tools come in: we used the R package metafor to fit a mixed-effect (ME) model to the data reported for each measure. In this way, we can estimate an overall interval of R2 values based on the effect sizes and the sample sizes. We can also estimate the interval of R2 that we can expect in future studies (this is called prediction interval). A compact way to represent these results is given by forest plots: for each study, we represent the effect size and the related sample size using a square and a horizontal error bar; then for each measure, we represent the results from the ME model using a diamond and an additional error bar; finally to represent the prediction interval we use two hourglasses and a dotted line.

The forest plot offers a detailed summary for each measure. What if we want to compare the R2 estimates across measures? To do that, we pooled together all the measures from all the studies and computed first a repeated measures meta-regression and then all the possible pairwise comparisons (Tukey's test), correcting for multiple comparisons (Bonferroni correction). To visually represent these results, we used two heatmaps, one for the z-scores and one for the p-values: each element refers to the comparison between the measure on the x axis and the one on the y axis.

Other factors to consider

Can some of this variance be explained by the differences in methodological choices and experimental conditions we mention? The number of studies is limited for a quantitative evaluation, but we can get a qualitative idea using bar plots and scatter plots organized by each condition.


Questions? Suggestions? Get in touch!