Emilie Poirion1, Daniel Garcia Lorenzo1, Isaac Adanyeguh1, Marie-Stéphane Aigrot1, Alexandra Petiet2, and Bruno Stankoff1,3
1Brain and Spine Institute, INSERM U1127/CNRS UMR 7225, Sorbonnes Universités, UPMC, CHU Pitié-Salpêtrière, 47 Bd de l'hôpital, 75013 Paris, Paris, France, 2Brain and Spine Institute, Center for Neuroimaging Research (CENIR), CHU Pitié-Salpêtrière, 47 Bd de l'hôpital, 75013 Paris, Paris, France, 3AP-HP, Saint Antoine Hospital, Department of Neurology, 184 Bd du Faubourg Saint Antoine, 75012 Paris, Paris, France
Synopsis
Experimental studies in mouse models
offer the opportunity to combine in-vivo
longitudinal high-field MRI and histological analyses. However, automatic MRI
tools for processing rodent data to avoid manual processing are lacking. We
proposed an automatic pipeline to perform systematic analyses on large murine cohorts
with longitudinal data. We first applied artifacts correction as bias
correction to optimize the subsequent steps. We then registered masks of regions
of interest (ROIs) for our analyses onto each subject from which we extracted
the quantitative data. This pipeline provides a way of quickly analyzing ROI regardless
of disease models or the MRI sequence.Purpose
In-vivo imaging of murine animal models in combination with
histological analysis is considered an essential tool that can shed light on
key processes underlying the physiopathology of complex diseases such as multiple
sclerosis (MS). Currently, tools for systematic analysis of MRI data in animals
remains a largely unmet need. So far, manual segmentation of regions
of interest (ROI) remains the most widely employed approach
1. In a longitudinal
study conducted on a group of mice treated with cuprizone – a model
characterized by the onset of widespread demyelination (which is the main
pathological feature of MS) followed by spontaneous remyelination upon
withdrawal of cuprizone – we developed a new tool to automatically segment ROIs
in mouse brain MRI data. We have thus generated an automatic pipeline, which
has the potential to be applied to longitudinal acquisitions, allowing a fully
automated analysis with minimal manual intervention.
Methods
Thirty-five C57B16 mice (8 weeks
old) received cuprizone mixed in food (0.2%) during twelve weeks to induce profound
demyelination. MRI scans were performed on an 11.7T system (Bruker Biospec
117/16USR) with a helium-cooled 1H quadrature transmit-receive
surface CryoProbeTM for mouse head. All mice underwent MRI scans
before and after intoxication at the 12th week, then 15, 30 and 45
days after the intoxication cycle. T2 relaxometry maps, which were
previously shown to reflect myelin content in this model, were obtained from a
multi-echo sequence (TR=5500ms, TE=15-120ms/15ms increments, resolution=100x100μm2,
matrix=128x128, slice thickness=200μm, Tacq=23mins).
We chose to analyze ROIs within
the corpus callosum (CC) and the putamen (Pu), as demyelination (Fig 1) and
remyelination are well described in these anatomical regions in the cuprizone mouse
model of MS2. Additionally, we chose to define ROIs in the
somatosensory cortex (S1), cerebellar peduncles (CbPed) and thalamic nucleus (Th).
To obtain reference measures, those ROIs were first manually drawn on 3
adjacent slices from the T2 maps on ParaVision5.1, and averaged to
get T2 values in those ROIs.
The automatic pipeline consisted
of four steps (Fig 2):
i) ROIs definition on the template: it was performed
manually using itksnap3 on a homemade template generated from a
group of wild-type healthy mice. This is done once per study.
ii) Bias correction: it was performed with a homemade tool based on Juntu et al4 from the image of the echo providing
the best contrast from the multi-echo sequence. These corrected images were
then used for the registration.
iii) Registration: timepoints within a subject were linearly
registered. From each subject, the first timepoint was non-linearly registered
onto the template using ANTS5.
iv) Values extractions: combination of the two
registrations allowed the transformation of ROIs back to each T2
map. The T2 value in each ROI was then extracted and averaged.
We first checked by visual inspection whether the automatically defined ROIs were consistent with anatomical regions and we excluded those that were considered
misregistered. Intra-class correlation (ICC) tests were used to assess the
accuracy of the T2 value from the automatic pipeline compared to the manual segmentation. We considered reproducibility to be good when ICC>0.7
and acceptable when 0.4<ICC≤0.7. We finally compared both the manual and the
automatic methods for processing duration.
Results
A hundred and twelve MRI scans
were analyzed. Exclusion percentage ranged from 18 to 35% (Fig 3) depending on
the region analyzed.
Results from ICC (Fig 3) showed good reproducibility at baseline for all ROIs (ICCCpu(J0)=0.93; ICCTh(J0)=0.97;
ICCS1(J0)=1.00; ICCCbPed(J0)=0.79) except for CC (ICCCC(J0)=0.68),
which was considered acceptable. CC kept acceptable reproducibility during
follow-up (ICCCC(J0)=0.68; ICCCC(J15)=0.62; ICCCC(J30)=0.62;
ICCCC(J45)=0.69). Other regions showed higher variability due to a
lower reproducibility observed at some timepoints: Cpu, ICCCpu(J0)=0.93
to ICCCpu(J30)=0.40; S1, ICCS1(J0)=1.00 to ICCS1(J30)=0.20.
In terms of processing duration,
manual segmentation required 112 hours while the automatic pipeline less than
24 hours, thanks to a queuing system.
Discussion
Our results showed good or acceptable reproducibility of the manual segmentation by the automatic pipeline. However, the choice of the ROI was critical for the success of the method. Particularly noisy regions due to the distance of the surface coil such as CbPed could not be accurately analyzed using the automatic pipeline.
Conclusion
We developed a new automatic pipeline to perform MRI analyses, which was faster, with minimal manual contribution from the operator. This pipeline can potentially be applied to measure demyelination/remyelination in cuprizone models with sensitive sequences such as T
2 mapping or diffusion-weighted imaging. This work should allow future
in-vivo longitudinal investigations of candidate molecules for promoting myelin repair in this model and other demyelinating diseases.
Acknowledgements
Program “Investissements d’avenir” ANR-10-IAIHU-06.
Ile-de-France Region (DIM Cerveau et Pensée)
Carnot Maturation grant 2014
References
1. Sun, S. W., Liang, H. F., Trinkaus, K., et al. Noninvasive detection of cuprizone induced axonal damage and demyelination in the mouse corpus callosum. Magnetic Resonance in Medicine, 2006; 55(2):302-308.
2. Skripuletz, T., Gudi, V.,
Hackstette, et al. De-and remyelination in the
CNS white and grey matter induced by cuprizone: the old, the new, and
the unexpected. Histol Histopathol, 2011:26(12):1585-1597.
3. Yushkevich, P. A., Piven,
J., Hazlett, H. C., et al. User-guided 3D active contour segmentation of anatomical
structures: significantly improved efficiency and reliability. Neuroimage, 2006:31(3):1116-1128.
4. Juntu, Jaber, Sijbers, J., Van Dyck, D., et al. Bias field correction for mri images. Computer Recognition Systems. Springer Berlin Heidelberg, 2005:543-551.
5. Avants, B. B., Tustison,
N. J., Song, G., et al. A
reproducible evaluation of ANTs similarity metric performance in brain
image registration. Neuroimage. 2011;54(3):2033-2044.