Valentin Demeusy1, Florent Roche1, Fabrice Vincent1, Jean-Pierre Guichard2, Jessica Lebenberg3,4, Eric Jouvent3,5, and Hugues Chabriat3,5
1Imaging Core Lab, Medpace, Lyon, France, 2Department of Neuroradiology, Hôpital Lariboisière, APHP, Paris, France, 3FHU NeuroVasc, INSERM U1141, Paris, France, 4Université de Paris, Paris, France, 5Departement of Neurology, Hôpital Lariboisière, APHP, Paris, France
Synopsis
We propose a novel automatic WMH segmentation method based on a convolutional neural network to study the longitudinal WMH changes among a cohort of 101 CADASIL patients. We demonstrate that this method is able to produce consistent quantitative measures of WMH volume by the strong correlation between the computed baseline WMH volume and the clinically assessed Fazekas score. Our main results show that the progression of WMH is correlated to the baseline volume and that this progression largely vary at individual level although a rapid extension is mainly detected between 40 and 60 years in the whole population.
Introduction
Cerebral
Autosomal Dominant Arteriopathy with Subcortical Infarcts and
Leukoencephalopathy (CADASIL) is the most frequent hereditary cerebral small
vessel disease (CSVD). The disease is caused by the accumulation of
extracellular-domains of the mutant NOTCH3 protein within the wall of cerebral
arterioles and capillaries. CADASIL is a severe condition that occurs during
mid-adulthood and can lead to stroke, mood disturbances, motor disability and
cognitive decline up to severe dementia. No preventative treatment is available
yet. MR imaging reveals in all affected individuals white-matter
hyperintensities (WMH) on T2-weighted or FLAIR images since the early stage of the
disease and developing with the disease progression. Quantification
and longitudinal tracking of WMH changes in CADASIL will be of
crucial importance to evaluate future disease treatments. The task is challenging due to major signal changes
observed in the cerebral tissue with aging with the disease.
Herein, we propose an innovative, accurate and reproducible convolutional
neural network (CNN) to track WMH along disease
progression.Methods
A sample of 101 CADASIL patients of age ranging from 24 to 74 years at
date of MRI exams was selected from a large cohort of CADASIL patients
evaluated over 15 years in the French National Referral center for Rare
Vascular diseases of the Brain and Retina (www.cervco.fr). During follow-up, we
monitored each patient using MRI from three to 10 times, every 18 months,
starting from the first exam (baseline). The MRI protocol included 3D T1
(1.0x1.0x0.8mm3) and 2D T2-weighted FLAIR (0.46x0.46x5.5mm3)
images. Expert clinicians and radiologists first segmented WMH using 2D FLAIR
images at baseline and scored WMH using the Fazekas1 scale. Thereafter, we developed an automatic WMH
segmentation method using a CNN 2,3 . In order to
train this network, we split our data into a training (32%), validation (13%)
and test
(55%) sets. Epoch and threshold
selection were performed by optimizing the results over the validation set. The
final performance of the model was assessed using the Dice score on the test
set only. When needed, the intracranial volume was obtained using ANTs 4 skull
strip and the whole brain segmentation by FreeSurfer 5.Results
The main characteristics of our patients from test dataset are
summarized in table 1. We first confirmed that the extent of WMH measured by the
Fazekas score was correlated to the age and WMH
volume assessed by our method but not to
the intracranial volume or the whole brain volume; correlations score are recorded in table 2. Our CNN model reached a mean Dice index of 0.84 over
the test (visual comparison on figure 3), while the reported inter-reader Dice index3 was
0.805. Previous work6 reported a Dice index of 0.80 using BIANCA7 in semi-automatic pipeline. Using our method, we found that the WMH volumes obtained at baseline were strongly correlated to the Fazekas scores (Spearman
r=0.921; p-value < 0.001) (c.f. figure 1). Our main
results showed a large variability of WMH changes at individual level with some
patients showing a large progression between 40 and 60 years of age while
others did not change even after 60 years. We found that the increase of WMH
during follow-up was larger in patients with Fazekas score between 3 and 5
(average progression of over 3.8mL/years) than in patients with Fazekas score
at 2 (average progression of 1.3mL/years). The results also showed a possible
plateau at early (Fazekas 1) and late (Fazekas 6) stages of the disorder where
the changes are the smallest (see figure 2).
Discussion
The results obtained
using automated measures of WMH appear promising. In the present study, our data showed that this innovative
approach allows obtaining measures correlated to semi-quantitative visual measures and that the progression of WMH along
follow-up was strongly correlated to these baseline measures as already and repeatedly
demonstrated in CSVD. But the results also suggest that
the progression of WMH largely vary at individual level further suggesting that we might have a group of fast and another of slow ‘progressors’ during
the course of the disease. There are limitations in this study; 1) the sample
of CADASIL patients remains relatively limited, 2) the
progression of WMH was not compared to a reference method, 3) the results were
obtained using 2D FLAIR images which are not well suited for volume assessment due
to their anisotropy and can lead to noisy intra-subject measures, 4) additional
studies are needed for evaluating potential correlations with clinical
manifestations, taking into account the number of lacunes,
microbleeds and degree of atrophy. In contrast, we can also consider that in
this study we included a large number of repeated MRI measures in patients with
a rare but archetypal CSVD, that the overall trends
were easily delineated and that our results appear consistent with what we
already learned from the natural history of CADASIL.Conclusion
Using a segmentation method based
on CNN, we obtained consistent quantitative
volumetric measures of WMH in CADASIL patients. Moreover, we measured their
longitudinal changes during short or long follow-up on repeated MRI
examinations. The results showed that the
progression of WMH can largely vary at individual level although a rapid
extension is mainly detected between 40 and 60 years in the whole population.Acknowledgements
This work was supported
by grant from the National Research Agency,
France (ANR-16-RHUS-0004 [RHU TRT_cSVD]).References
- Fazekas, F., Barkhof, F., Wahlund, L., Pantoni, L., Erkinjuntti, T., Scheltens, P., & Schmidt, R. (2002). CT and MRI Rating of White Matter Lesions. Cerebrovasc Dis, 31-36. doi:10.1159/000049147
- Szegedy, C., Ioffe, S., Vanhoucke, V., & Alemi, A. (n.d.). Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning. Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence.
- Ghafoorian, M., Karssemeijer, N., Heskes, T., van Uden, I. W., Sanchez, C. I., Litjens, G., . . . Platel, B. (2017). Location Sensitive Deep Convolutional Neural Networks for Segmentation of White Matter Hyperintensities. Nature.
- Avants, B., Tustison, N., &
Song, G. (2009). Advanced normalization
tools (ANTS). Insight Journal.
- Fischl, B. (2012). FreeSurfer. NeuroImage.
-
Ling, Y., Jouvent, E., Cousyn, L., Chabriat, H., & De Guio, F. (2018). Validation and Optimization of BIANCA for the Segmentation of Extensive White Matter Hyperintensities. Neuroinformatics.
- Griffanti, L., Zamboni, G., Khan, A., Li, L.,
Bonifacio, G., Sunderesan, V., . . . Jenkinson, M. (2016). BIANCA (Brain
Intensity AbNormality Classification Algorithm): A new tool for automated segmentation
of white matter hyperintensities. Neuroimage.