SEGMENTATION OF ANATOMO-FUNCTIONAL 3D POST MORTEM DATA USING A MRI-BASED 3D DIGITAL ATLAS: TRANSGENIC MOUSE BRAINS STUDY J. Lebenberg1 , A-S. H´ erard1, A. Dubois1, M. Dhenain1, P. Hantraye1 , V. Frouin2 and T. Delzescaux1 1

MIRCen-LMN-CNRS-URA 2210, CEA-DSV-I2BM, Fontenay-aux-Roses, France 2 SCSR, CEA-DSV-I2BM, Evry, France ABSTRACT

Many analyses in neurosciences are carried out on histological and autoradiographic datasets and performed by manually drawing regions of interest on these 2D post mortem data. Such task being time-consuming, we propose an automated segmentation strategy to analyze 3D post mortem brain images. This method is based on the co-registration of a MRIbased 3D digital atlas on 3D-reconstructed post mortem data. We first deformed the original MRI on post mortem volumes. We then applied deformation parameters to warp the digital atlas. The method was tested on APP/PS1 mouse models of Alzheimer’s disease and PS1 control littermates. Its reliability was qualitatively evaluated and a quantitative assessment was performed by comparing atlas and manual segmentations. Our results show this approach is promising to investigate at organ scale post mortem mouse datasets faster and in a more robust manner than with classical methods. Index Terms— Biological image processing, Image registration, Image region analysis, Mouse brain atlas 1. INTRODUCTION Murine models of neurodegenerative diseases, like Alzheimer’s Disease (AD), are commonly used to improve pathology knowledge and to test drug effects [1]. Histological staining and ex vivo autoradiography remain the gold standards for the accurate description of neuroanatomy and brain function ([1],[2]). Such post mortem data are classically analyzed by manually outlining regions of interest (ROI) on a limited number of sections. However, this step is laborintensive, time-consuming and subject to operator bias. To overcome these limitations, we propose to reconstruct in 3D post mortem data and to analyse them by using a digital atlas segmentation approach. Several 3D digital rodent brain atlases derived from MR images have recently become available ([3] , [4]). We thus propose a strategy to register 3D MRI atlas on post mortem data in order to fully and automatically segment cerebral structures on post mortem datasets. The reliability of the method was assessed qualitatively and quantitatively by comparing the atlas-based segmentation

with manual segmentation, considered as gold standard. The method was applied on transgenic mice modelling AD and their control littermates.

2. MATERIAL 2.1. MRI-based 3D digital atlas The MRI-based 3D digital atlas used for our study was downloaded on the Center for In Vivo Microscopy website (http://www.civm.duhs.duke.edu/). It derived from T1 and T2-weighted 3D MR images of six 9-week-old C57BL/6J (wild type) mice recorded after brain active staining [3]. The isotropic scans resolutions are 21.5µm and 43µm respectively. The images were acquired within the cranium vault to avoid distortions due to brain extraction. Thirty three anatomical structures were segmented, among which the cerebral cortex, the hippocampus and the striatum.

2.2. Animals and Biological data Our method was applied to 4 APP/PS1 mouse model of AD (right hemisphere, 64±1 week-old) and 3 PS1 control (ctrl) animals (right hemisphere, 65±2 week-old) [5]. The genetic background of these transgenic mice was C57BL/6J. All the procedures were carried out in accordance with the recommendations of the EEC (86/609/EEC) and the French National Committee (87/848) for the use of laboratory animals. [14 C]-2-deoxyglucose was injected in vivo (16.5µCi/100g body weight) to evaluate cerebral glucose uptake by autoradiography. Brains were cut on a cryostat into 20µm-thick coronal serial sections (olfactory bulb and cerebellum were excluded). Serial sections (one out four) were mounted on glass slides and exposed onto autoradiographic film. The same sections were then processed for Nissl staining to provide complementary information about brain anatomy. Images from the surface of the brain was recorded by using a digital camera with in-plane resolution of 27×27µm2.

3. METHODS 3.1. 3D reconstrution of post mortem images Autoradiographs, with [14 C] standards, and histological sections were digitized using a flatbed scanner (8 bits, 1200 dpi, pixel size 21x21 µm2 ). As described in [6], these post mortem data were individualized and then stacked thanks to BrainRAT, a new add-on of BrainVISA (free software, http://brainvisa.info/). Blockface photographs captured prior to each section allowed spatially consistent 3D volume reconstruction, as each histological section was first registered with the corresponding blockface photograph and each autoradiographic section was thereafter registered with the corresponding registered histological section. To measure glucose uptake, the gray level intensities of autoradiographic volumes were calibrated using the co-exposed [14 C] standard scale and thus converted into activity values (nCi/g). Corrective coefficients were applied to normalize brain activity so as to allow comparison between strains [2]. 3.2. MRI-based atlas and post mortem data registration strategy To co-register the atlas with post mortem data, the T1weighted MR image was first deformed to match with each 3D-reconstructed post mortem volumes of our dataset. The registration strategy consisted in gradually increasing degrees of freedom to register images first globally, then locally. It was defined according to the following steps: 1. A global rigid transformation using the mutual information as similarity criterion was first estimated. 2. An affine deformation initialized with previously computed parameters was then calculated with the Block Matching registration technique [7] (optimization based on correlation coefficient). 3. Finally, to locally enhance MRI and post mortem data registration, we deformed this image using an elastic transformation (Free Form Deformation) initialized with the previous transformation (optimization based on mutual information) [8]. Putting 10 × 10 × 10 control points regularly spread out the volume, this deformation used 3000 degrees of freedom.

were measured by computing the Dice coefficient defined as: VA ∩ VM (1) VA + VM where VA and VM are the volumes of the atlas-based and manual segmentations, respectively. According to the literature, a Dice coefficient greater than 0.7 is considered as a good agreement [9]. Thus, to validate the atlas registration, hippocampus (Hc), cortex (Cx) and striatum (Striat) were manually delineated on the histological volume of one APP/PS1 and one PS1 mouse by an expert. These ROIs were chosen because of their different localization in mouse brain and their different size: Cx is a large external structure spread all over the brain whereas Hc and Striat are smaller inner structures mainly localized in the posterior and anterior parts of the brain, respectively. In addition to Dice coefficients computation, we compared ROI mean activity measured with both segmentations (µact(A) and µact(M) being atlas-based and manual segmentations, respectively). Variation coefficients (δµ ) were computed for each structure to estimate atlas segmentation error in comparison with the gold standard measures (manual segmentations): κ=2×

δµ =

|µact(A) − µact(M) | µact(M)

(2)

3.4. Segmentation of anatomo-functional data Our methodology was then applied onto the dataset to obtain anatomo-functional parameters. Supplementary ROIs were studied: the inferior colliculus (IC), the superior colliculus (SC) and the thalamus (Thal). A two-sample t-test was then computed to compare both strains. 4. RESULTS 4.1. Data reconstruction Figure 1 presents T1-weighted MR image with its corresponding MRI-based 3D digital atlas (left part) and 3Dreconstructed post mortem images from one APP/PS1 (right part). We cropped the MR images and atlas according to the dashed rectangle to obtain a brain volume similar to our data. Registration strategy was applied on these areas.

For each step, we tested the three spatially consistent post mortem volumes as reference image for the registration. The deformation parameters were then used to warp the 3D digital atlas in the post mortem geometry. 3.3. Registration evaluation Registration accuracy was qualitatively assessed by visual inspection of the superimposition of the inner and outer contours of MR image on post mortem data. To quantitatively evaluate the registration strategy, agreements between atlas-based and manually delineated ROIs

Fig. 1. T1-weighted MRI with its corresponding MRI-based 3D atlas (left part) and 3D-reconstructed post mortem data (right part).

4.2. Registration validation For each step of the registration strategy, we chose a reference image according to our qualitative evaluation and Dice coefficients computations. According to results, we assumed that the chosen reference modality for registration steps did not have any influence on the registration quality. Detailled results will be shown in a future work. The blockface volume was thus chosen for all registration steps because of its highest spatial consistency as regarded to the other modalities. Figure 2 and Table 1 sum up obtained results. Figure 2 shows, in coronal view, the superimposition of the contours of the T1-weighted MR image on a APP/PS1 3Dreconstructed histological volume (Fig. 2A) and after each step of the registration strategy: rigid registration (Fig. 2B), affine registration (Fig. 2C) and elastic registration (Fig. 2D). Rigid transformation was able to center both images (Arrow 1 between figures 2A and 2B). Affine transformation was able to compensate volume differences between atlas and experimental data (Arrow 1 between figures 2B and 2C). Finally, elastic transformation allowed to locally adjust the registration between atlas and experimental data: in figure 2D, the external contours (1) and those defining inner structures such as the corpus callosum (2) or the hippocampus (3) are correctly superimposed. 22 1 33 1

22 33

22

22

3 33

3 1

A) A)

B) B)

1

1

C) C)

11

D) D)

1

Fig. 2. Superimposition of T1-weighted MR image contours on a APP/PS1 histological volume before (A) and after rigid (B), affine (C) and elastic (D) registrations. Arrows 1, 2, 3 indicate external contours, corpus callosum and hippocampus, respectively.

(blue). Similar visual inspections were performed on all manually segmented ROIs (cortex and striata). ROI

A visual understanding of the atlas registration on one APP/PS1 experimental data is showed in Figure 3. This figure presents the studied part of the 3D digital atlas (Fig. 3A) and shows that even if atlas and experimental volumes were not initially aligned (Fig. 3B), atlas outer edges (light red) finally fit with those of the blockface volume (light blue) after complete registration (Fig. 3C). Moreover, the atlasbased segmentation of the hippocampus (red) presented, once registered, a good matching with the manual segmentation

κ rig

κ aff

κ elast

0.79 0.79 0.81

0.83 0.83 0.84

A) PS1 mouse (ctrl) Cx Hc Striat

0.45 0.38 0.45

0.76 0.82 0.80

B) APP/PS1 mouse (AD model) Cx Hc Striat

0.56 0.41 0.47

0.79 0.82 0.80

0.81 0.82 0.81

0.85 0.86 0.82

Table 1. Dice coefficient (κ) computed for cerebral cortex (Cx), hippocampus (Hc) and striatum (Striat) for one PS1 (A) and one APP/PS1 (B) mouse before (init) and after each step of the registration strategy (rig, aff and elast).

A)

B)

C)

Fig. 3. Studied part of mouse brain (dashed line) in (A). Superimposition of a APP/PS1 blockface volume with its Hc manually segmented (blue) and the atlas with its segmented Hc (red) before (B) and after (C) registration.

Table 2 compares µact measured in Cx, Hc and Striat of one PS1 (Tab. 2A) and one APP/PS1 (Tab. 2B) mouse with manual and atlas segmentations. Whatever studied ROI, computed variation coefficients (δµ ) show that differences between µact(M) and µact(A) were minor (δµ ≤ 5%). ROI

Table 1 displays Dice coefficients computed for Cx, Hc and Striat of one PS1 (Tab. 1A) and one APP/PS1 (Tab. 1B) mouse before (init), after rigid (rig), affine (aff) and elastic (elast) registration. Globally, for each ROI, this criterion varies in a similar way whatever the mouse. Data centering observed after rigid transformation in Fig. 2B) is illustrated by the most important gain of Dice index after this step (∼80%). Registration enhancement is then shown by κ continuous increase.

κ init

µact(M ) (nCi/g)

µact(A) (nCi/g)

δµ

265.53 ± 48.36 239.31 ± 41.88 273.21 ± 33.59

< 0.01 0.01 0.02

A) PS1 mouse (ctrl) Cx Hc Striat

265.38 ± 49.57 240.61 ± 40.49 268.25 ± 37.68

B) APP/PS1 mouse (AD model) Cx Hc Striat

275.90 ± 61.20 225.89 ± 38.61 264.56 ± 48.83

279.37 ± 57.608 225.52 ± 40.06 276.19 ± 40.87

0.01 < 0.01 0.04

Table 2. Cerebral cortex (Cx), hippocampus (Hc) and striatum (Striat) mean activities for one PS1 and one APP/PS1 mouse measured with manual and atlas segmentations (µact(M ) and µact(A) ).

The effectiveness of the atlas-based segmentation on our experimental data is demonstrated by final Dice coefficients higher than 0.7 and minimal variation coefficients between both µact (δµ ≤ 5%). This atlas could automatically segment data instead of an expert.

4.3. Group study We performed 3D reconstruction on all post mortem data of our study and registered the MRI on each subject. Similar visual inspections as those presented in figure 2 were reproduced on the 7 mice of this study. Once the MRI-based atlas registered, we computed average structure volumes (V ) and mean structure activities (µact ). Results are presented in Table 3. ROI

PS1 mice (n = 3)

APP/PS1 mice (n = 4)

3

A) Volume V (mm ) Right hemisphere Cx Hc IC SC Striat Thal

176.77 ± 14.18 75.64 ± 7.29 11.89 ± 0.55 2.48 ± 0.26 6.12 ± 0.73 12.53 ± 0.28 17.27 ± 1.32

175.57 ± 3.92 75.28 ± 1.30 12.9 ± 0.48 2.49 ± 0.23 5.55 ± 0.11 12.58 ± 0.50 16.49 ± 0.75

the difficulty to compare segmentations based on images acquired within and without cranium and with different imaging modalities. Complementary tests will be realised to validate our method. Finally, comparing both strains, on the scale of ROI, did not lead to statistically significant differences (morphometric or functional). These results are in agreement with previous studies [10]. The present study confirms that our methodology was able to successfully co-register MRI from a wild type mice with 3Dreconstructed post mortem brains from two different transgenic strains. The MRI-based atlas is well-fitting to our study and could be used as a template for fully-automated mouse brain segmentation. This method is easier and faster than a manual analysis. It is also more objective because nonoperator dependant. This approach is promising to investigate, on the scale of organ, a larger number of post mortem volumes and could find several applications in exploratory studies in neuroscience.

B) Activity µact (nCi/g) Right hemisphere Cx Hc IC SC Striat Thal

216.04 ± 5.4 270.54 ± 4.64 237.67 ± 1.86 233.11 ± 28.67 252.78 ± 12.08 284.04 ± 9.62 263.76 ± 9.4

222.51 ± 3.44 272.68 ± 4.77 228.24 ± 6.79 295.80 ± 53.96 258.68 ± 13.02 274.3 ± 1.74 242.96 ± 3.56

Table 3. V and µact per strain (PS1 versus APP/PS1) measured with atlas segmentation.

No statistically significant volumetric nor functional differences were found between groups on the scale of ROI. 5. DISCUSSION AND CONCLUSION The 3D reconstruction of post mortem data strategy provided 3 spatially consistent post mortem volumes for each subject. Our registration strategy, which gradually increased degrees of freedom applied on the MRI to match with post mortem volumes, allowed to register images with a coarse-to-fine approach. This method also enabled algorithms to speed up optimization process. The choice of the blockface volume as reference image facilitates registration thanks to its highest spatial consistency as regards to the other modalities and its likeness with the MRI. Superimposition of the contours of the MRI on 3D histological volume, as well as superimposition of atlas-based segmentations on manual ones showed that MRI and derived atlas could be progressively deformed to match with post mortem data. Final Dice coefficients higher than 0.7 and minor µact variation coefficients attested to the similarity of results obtained with both segmentation methods. Errors in segmentation could be due to the misregistration but also demonstrated

6. REFERENCES [1] P.C. Wong et al., “Genetically engineered mouse models of neurodegenerative diseases.,” Nat Neurosci, vol. 5, no. 7, pp. 633–639, 2002. [2] J. Valla et al., “Age- and transgene-related changes in regional cerebral metabolism in psapp mice.,” Brain Res, vol. 1116, no. 1, pp. 194–200, 2006. [3] G.A. Johnson et al., “High-throughput morphologic phenotyping of the mouse brain with magnetic resonance histology.,” Neuroimage, vol. 37, no. 1, pp. 82–89, 2007. [4] A.E. Dorr et al., “High resolution three-dimensional brain atlas using an average magnetic resonance image of 40 adult c57bl/6j mice.,” Neuroimage, 2008. [5] B. Delatour et al., “In vivo mri and histological evaluation of brain atrophy in app/ps1 transgenic mice.,” Neurobiol Aging, vol. 27, no. 6, pp. 835–847, 2006. [6] A. Dubois et al., “BrainRAT: Brain Reconstruction and Analysis Toolbox. a freely available toolbox for the 3D reconstruction of anatomo-functional brain sections in rodents,” 38th Annual Meeting of the SFN, 2008. [7] S. Ourselin et al., “Reconstructing a 3D structure from serial histological sections,” Image and Vision Computing, vol. 19, no. 1-2, pp. 25–31, 2001. [8] D. Mattes et al., “Pet-ct image registration in the chest using free-form deformations,” IEEE Trans Med Imaging, vol. 22, 2003. [9] S. Maheswaran et al., “Analysis of serial magnetic resonance images of mouse brains using image registration.,” Neuroimage, vol. 44, no. 3, pp. 692–700, 2009. [10] M. Sadowski et al., “Amyloid-β deposition is associated with decreseased hippocampal glucose metabolism and spatial memory impairment in app/ps1 mice,” Journal of Neuropathology and Experimental Neurology, vol. 63, no. 5, pp. 418–428, 2004.