Introduction

Establishing an individual prognosis in heterogeneous neurological diseases such as multiple sclerosis is a challenging task for neurologists. Moreover, in multiple sclerosis, clinical disability correlates poorly with the classical radiological measures based on lesion load (e.g. lesion count), yielding the so-called ‘clinico-radiological paradox’.
Previous studies showed the clinical relevance of characterizing lesion location within the major white matter tracts using an atlas-based approach to overcome the limitations that come with the absence of diffusion imaging in clinical MR protocols¹ for multiple sclerosis. Here, we aim at modelling the affected structural brain connectivity using brain graphs and assess the clinical relevance of such models to improve the prediction of future disease progression in multiple sclerosis patients.

Dataset

A prospective longitudinal observational study enrolled patients within four months after their first clinical event suggestive of multiple sclerosis². A subset of 107 patients with complete data and treated with interferon-beta over the whole study duration (40 males, age=30.2±7.7 years) was retained for analysis. MRI scans were acquired at 1.5T (Philips Gyroscan NT 15, Best, The Netherlands) and included fluid attenuated inversion recovery (FLAIR) with 1.5 mm thickness (TR/TE/TI 11000/140/2600 ms) and 3D spoiled-gradient recalled (SPGR) images with 1 mm slice thickness (TR/TE 25/5 ms). Hyperintense lesions were manually segmented in FLAIR.

Methods

Graph Model of Disconnectome
A tractography atlas generated from 842 healthy subjects³ was used together with the Brainnetome parcellation atlas⁴ to study structural connectivity. To extract the affected streamlines for each patient, the lesion mask was spatially registered and superimposed to the atlas tractogram in MNI space. Then, only streamlines passing through lesions were retained (see Figure 1A). The disconnectome was modelled as a brain graph $$$G_{dis}$$$ whose vertices $$$V$$$ represented brain regions and whose edges $$$E_{dis}(i,j)$$$ represented the number of streamlines connecting two areas $$$i$$$ and $$$j$$$ (see Figure 1B). Similarly, the atlas tractogram was modelled as a brain graph $$$G_{atlas}$$$ with vertices $$$V$$$ and edges $$$E_{atlas}(i,j)$$$ representing the number of streamlines connecting two regions in an average healthy subject.

Each vertex of $$$G_{dis}$$$ was weighted by the relative percentage of affected streamlines computed as:
$$RAS(i) = \frac{\sum_{j \in V_i}E_{dis}(i,j)}{\sum_{j \in V_i}E_{atlas}(i,j)}$$
with $$$V_i$$$ the set of neighbours of vertex $$$i$$$.

Finally, a graph of preserved connectivity $$$G_c$$$ was created by combining $$$G_{dis}$$$ and $$$G_{atlas}$$$ such as:
$$E_c(i,j)=\begin{cases}\epsilon,&\text{if }E_{atlas}(i,j)=0\\ \frac{E_{atlas}(i,j)-E_{dis}(i,j)}{E_{atlas}(i,j)}+\epsilon,&\text{otherwise} \end{cases}$$
with $$$\epsilon=10^{-5}$$$ to ensure a connected graph.

For representation purposes, the parcellation areas were grouped into the main brain lobes, and the overall inter- and intra-lobes affected connections were computed. An intuitive visual representation of an individual’s disconnectome graph was created and is shown in Figure 1C.

Disease Progression Modelling
We studied the contribution of graph-based features to the definition of disease phenotypes with diverse prognosis. We built and compared models based on 1) the RAS metric and 2) the average clustering coefficient of the whole graph to baseline models including 3) lesion volume and 4) lesion count. For each model, patients were first separated into two clusters using a k-means algorithm on each of the four lesion representations. Then, we fitted survival models for each cluster predicting the advent of a relapse and compared the Kaplan-Meier curves using a log-rank test.

Results

The relative percentage of affected streamlines per vertex was plotted for every patient in Figure 2 together with the lesion load and the time duration between the baseline visit and the first relapse. The heterogeneity of the data does not allow to identify a consistent pattern of disconnected brain areas nor lesion load metrics when related to the time to first relapse. However, left parietal, insular and subcortical regions seem to be generally more disconnected than other brain areas across subjects.

Figure 3 shows the outcome of k-means clustering and survival analysis for the different models. By looking at the risk tables, one can assess that cluster sizes are comparable between models. Clustering based on lesion count doesn’t result in two phenotypes with different disease courses ($$$\chi^2 = 1.7, p=0.2$$$), whereas other models allow a significant discrimination between the Kaplan-Meier curves (lesion volume: $$$\chi^2 = 4.9 , p = 0.027$$$ ; RAS: $$$\chi^2 = 5.5, p = 0.019$$$; average clustering coefficient: $$$\chi^2 = 6.2 , p = 0.013$$$).

Discussion/Conclusion

We proposed an approach to quantitatively characterize the affected structural connectivity in multiple sclerosis patients without requiring diffusion imaging. Graph features were shown to allow the definition of two distinct disease phenotypes characterized by different disease courses, outperforming the baseline model based on simple lesion count and performing at least as well as lesion volume, while offering additional insights into spatial lesion distribution.

Because our method relies on spatial registration of patients’ brain to standard healthy templates, it is important to restrict such analysis to young patients at early stages of the disease to avoid registration limitations that would be induced by brain atrophy and aging.

In addition to providing potentially useful information to neurologists for the prediction of individual prognosis, we expect that our disconnectome representation could contribute to a better understanding of (dis-)connection mechanisms underlying various neurological disorders. Notably, our analysis could be applied retrospectively on previously acquired clinical datasets as standard clinical MR sequences are sufficient.

Acknowledgements

No acknowledgement found.