LATEST TRENDS on APPLIED MATHEMATICS, SIMULATION, MODELLING

Automated Workflow for Spatial Alignment of Multimodal MR Image Acquisitions in a Longitudinal Study of Cognitive Aging Erlend Hodneland, Martin Ystad, Judit Haász, Antonella Zanna Munthe-Kaas and Arvid Lundervold

Abstract—In this application-oriented investigation we describe a framework and the challenges of registering multimodal brain MR images from a cohort of more than hundred subjects. Image examinations are done three years apart and consist of 3D highresolution anatomical images, low-resolution tensor-valued DTI recordings and 4D resting state fMRI time series. The registration procedures are incorporated in multi-subject statistical analyses, combining image-derived information with cognitive test results and genotypes. Due to the large number of subjects an automated and time-efficient workflow (e.g. scripting) is strongly desired, putting constraints on the registration methods. Keywords—Magnetic resonance imaging of the brain, Diffusion tensor imaging, Functional MRI, Image registration, Image segmentation, Automation. I. INTRODUCTION

T

HE present work is related to a longitudinal study of cognitive aging following a cohort of more than 100 healthy elderly people. Two waves, about three years apart, have been completed in the ongoing study, each consisting of extensive neuropsychological assessment and multimodal MRI examinations. Genetic profiling was performed in the first wave. A major aim of the study is to investigate how variations in neuropsychological test scores (cognitive aging) correlate with changes in MRI parameters (e.g. brain volumes, cortical

Manuscript received June 30, 2010. This work was supported in part by the MedViz consortium (Project 7) and Helse Vest grant #911593. E. Hodneland is with the Department of Biomedicine, Neuroinformatics and Image Analysis Laboratory, and Department of Mathematics, University of Bergen, Bergen, Norway (e-mail: [email protected]). M. Ystad is with the Department of Biomedicine, Neuroinformatics and Image Analysis Laboratory, University of Bergen, Bergen, Norway (e-mail: [email protected]). J. Haász is with the Department of Biomedicine, Neuroinformatics and Image Analysis Laboratory, University of Bergen, Bergen, Norway and Department of Neurology, Haukeland University Hospital, Bergen, Norway (e-mail: [email protected]) A.Z. Munthe-Kaas is with the Department of Mathematics, University of Bergen, Bergen, Norway (e-mail: [email protected]) A. Lundervold (corresponding author) is with the Department of Biomedicine, Neuroinformatics and Image Analysis Laboratory, University of Bergen, Norway and Department of Radiology, Haukeland University Hospital, Bergen, Norway (phone: +47 55586353; fax: +47 55586360; email: [email protected]).

ISSN: 1792-4332

264

thicknesses, and white matter integrity) over time. In the second wave a “resting state” and a “finger-tapping” fMRI protocol (4D data) was added to the 3D anatomical scans and diffusion tensor imaging (DTI) scan. This was added in order to assess resting state networks (RSNs) in the brain, including the “default mode network” (DMN) [1], and functional and structural connectivity between segmented brain regions [2]. In this comprehensive neuroimaging context we present image registration challenges related to (i) intra-subject, multi-modal MR image registration between 3D high-resolution-anatomical images, low resolution tensor-valued DTI recordings, and 4D resting state fMRI; (ii) intra-subject registration between image recordings obtained at wave-1 and wave-2, respectively; and (iii) inter-subject registration for group analysis (e.g. group ICA). In addition, evaluation of registration accuracy and workflow performance are important issues in this multi-subject, multi-modal application. In the following, we will mainly focus on topic (i) and workflow requirements.

II. METHODS A. Multimodal MRI acquisitions Some spatial characteristics of the multimodal MRI protocol being used are given in Table 1. Note the various grid and voxel sizes between the different modalities (MR measurement techniques), and between original acquisitions and processed data e.g. FreeSurfer-segmented anatomy, fractional anisotropy (FA) calculated from eigen-decomposition of diffusion tensor images, and spatial ICs derived from independent component analysis (ICA). TABLE I THE MULTIMODAL MRI PROTOCOL FROM WAVE-1 AND WAVE-2 Modalit y

T1-w anatomy DTI fMRI wave-2 only

Dim i × j × k × t

Voxel size (mm3)

Original acq. FreeSurfer

256×256×124 256×256×256

Original acq. FA image Original acq. ICA components

256×256×25×30 256×256×25 64×64×25×256

0.94×0.94×1.4 1.0×1.0×1.0 0.94×0.94×4.0 0.94×0.94×4.0 3.75×3.75×5.5 3.0×3.0×3.0

Image type

53×63×46

Examples of recorded and calculated images are shown in Fig.

ISBN: 978-960-474-210-3

LATEST TRENDS on APPLIED MATHEMATICS, SIMULATION, MODELLING

1 (note that all images shown are in original matrix orientation). The T1-weighted anatomical image (Fig.1 B) is used for skull stripping, tissue segmentation and brain surface reconstruction by FreeSurfer (http://surfer.nmr.mgh.harvard.edu) [3]–[5]. In this process a re-sampled image with isotropic voxels is created, shown in Fig.1 E. The DTI recordings (Fig.1 A) reflect voxel-wise degree and direction preferences of water diffusion in the brain. These local water diffusion measurements are strongly correlated with the anisotropic microstructure. In white matter (= myelinated nerve fibers) the derived FA value is often interpreted as a measure of “white matter integrity” [6] or “structural brain connectivity” that might deteriorate in aging [7]. An example of a calculated FA image (FA map) is shown in Fig.1 D). Eigen decomposition of the tensor map, derived from the DTI recordings, can also be used to calculate fiber connections in the brain (i.e. fibertracking). The fMRI time series (Fig.1 C) reflect local neuronal activity through the blood oxygen level dependent (BOLD) contrast mechanism. From the 4D resting state fMRI data we performed an independent component analysis (ICA) to detect spatially independent resting state networks in the aging brain [8], [9] – the ICA components (Fig.1 F). B. Comment on manual interaction vs automated workflow In our application, an automated and time-efficient workflow is strongly desired. Fully automated approaches to image registration and segmentation have several advantages compared to methods that require manual interaction, e.g. (i) automation enable more easily repeated evaluations for different parameter settings; (ii) results are reproducible and less subjective; (iii) misalignment (or errors) are often systematic and more easy to detect at early stages, whereas, in manual approaches, errors are normally less predictable and coalesce with the data in non-suspicious way, hence harder to detect. However, one should ensure that the accuracy of the automated method is adequate for the particular application, and robust to typical variations in data quality.

In some cases a semi-automated approach is necessary to achieve these goals. In our application, we needed to correct FreeSurfer segmentations manually in almost every subject, using control points in white matter, since the automated segmented regions turned out to be unacceptable or impossible according to anatomical knowledge. This was the only manual intervention needed in our processing chain. C. Multimodal images and registration As seen from Table 1, images are of different dimension, orientation and voxel size, and they are also misaligned. Therefore, the images cannot be compared directly without spatial alignment. Nonlinear registration typically uses an affine registration as initialization and then a further deformation field is computed to account for local, nonlinear effects. Between images within the same modality, both affine and nonlinear registration algorithms normally have a good performance [10]–[12]. See also Fig.5. There are several reports on successful nonlinear intermodality registration [13]–[16], and in some cases it is better to perform a nonlinear registration into established template spaces for each modality separately [13], [16], [17], [18], [19]. However, for our data of different image modalities, we found an affine registration (implemented in FSL) more reliable than a nonlinear approach. The reason for this is probably the lack of simultaneous information between the images of different modalities. As we are dealing with a large number of subjects, it is desirable to avoid time-consuming validation of the results that occasionally might be substantially impaired by errors. A further argument for applying affine registration is that brain images typically exhibit less nonlinear changes than images from other organs, mainly due to the fixed skull cavity. Therefore, we applied affine registration to our data. The data should normally be registered to the image of the highest resolution since the template should have the largest amount of information. Also, a trial registration from the highresolution anatomy domain into the low-resolution DTI domain resulted in some empty anatomy regions due to interpolation effects. Due to these reasoning and observations, we registered the fMRI data and DTI data into the anatomy domain. D. DTI and intra-modal eddy current correction The time-varying magnetic fields in MRI scans, such as DTI with multiple acquisitions, can induce eddy currents in the brain. These currents are altering the local magnetic field and will thus create geometric distortions in the image that might cause problems in further analysis [20]. To counteract these undesired effects, we applied the tool (eddy correct) in the FSL package (http://www.fmrib.ox.ac.uk/fsl) for eddy current correction. This shell-script employs affine registration between each of the 25 diffusion sensitive (b=1000) and

Fig. 1 Examples of recorded and calculated, unregistered images. The original image acquisitions (first row) are used to produce the processed images (second row). The images in the different modality “domains” are not aligned in image space, and can therefore not be compared directly without a

ISSN: 1792-4332

proper registration. The registration is performed between the processed images, from image D to E and from image F to E, indicated by arrows. The dimensions (grid size) of the images are in each case shown in the lower left corner.

265

ISBN: 978-960-474-210-3

LATEST TRENDS on APPLIED MATHEMATICS, SIMULATION, MODELLING

direction selective DWI images and the 5 baseline (b=0) images to a fixed reference volume using FSL’s flirt. Figure 2 shows the skull-stripped (FSL’s bet) FA-image without (A) and with (B) eddy current correction. The difference image (C) reveals the presence of edge structures, due to the effect of minor geometric distortions between the combined and uncorrected DWI images. This will influence the statistical distribution of FA values in certain white matter regions and thus the correlations between cognitive test scores and “white matter integrity”.

Fig. 2 Eddy current correction of the axial DTI data. The FA image without (A) and with (B) eddy current correction and their difference image (C).

E. Registration of DTI data to the high-resolution anatomy An affine registration between the calculated FA image and the high-resolution 3D anatomical image usually works well for most of the subjects in our cohort (cf. Fig. 3). The eddy current corrected FA image (A2) was here registered (by flirt) to the isotropic FreeSurfer image (B2). The resulting registered image is shown in C1 after applying the obtained spatial transformation to the none-skullstripped (unbet’ed) FA image (A1). Upon visual inspection of the continuity of edges inside the brain, representing brain structure boundaries across “white” and “black” squares in the checkerboard image (D1), it is evident that the registration quality is good. However, for some of the subjects this affine and “blind” registration (i.e. not using DICOM header information about voxel positions in scanner coordinate system) failed completely. We therefore parsed and extracted the DICOM header information into the NIFTI’s sform and qfrom quaternions, providing image voxel positions in “laboratory frame” for each of the images being registered. This was meaningful since the multimodal MRI acquisitions were performed during the same imaging session, without repositioning of the subject. This header information is very useful since it provides a high-quality initial approximation to the spatial transformation. Again, we used FSL’s flirt to find an initial affine transformation from header information in order to register the FA image (A1) to the FreeSurfer anatomical image (B1). Thereafter, a second affine registration was applied between A2 and B2 in case the subject had moved slightly. The resulting transformed image is shown in Fig. 3 (C2) with the corresponding checkerboard image (D2) to the template.

ISSN: 1792-4332

266

Fig. 3 Inter-modality registration of DTI (FA image) to the

FreeSurfer

anatomical image. The registration was tested for three different methods, an affine “header-blind” registration (C1 and D1), affine “header-based” registration (C2 and D2) and affine with nonlinear registration (C3 and D3). The resulting transformed FA images are shown in row C and the checkerboard images with the templates are given in row D. From visual inspection of the checkerboard-images it is clear that the affine registrations are more successful than the nonlinear approach.

F. Multimodal registration of DTI white matter tracts and fMRI resting state networks into 3D MRI anatomy An important goal in the “cognitive aging” project is to assess structural and functional connectivity between specific brain regions defined partly anatomically (FreeSurfer segmentation) and partly functionally (resting state networks obtain by spatial ICA). Moreover, such brain connectivity assessments are performed subject-by-subject and compared to scores on neuropsychological tests, and also group-wise (e.g. those subjects with at least one APO4ε allele in one group, those who are APO4ε-negative in the other group, where APO4ε is a vulnerability gene for Alzheimers disease). Registration of DTI-derived fiber tracts and fMRI resting state network (“blobs”) in 3D anatomical space for one subject is illustrated in Fig. 4, using the same affine transformation as for FA image to FreeSurfer anatomy. The figure demonstrates the great potential of image registration and segmentation methodology in neurocognitive research.

ISBN: 978-960-474-210-3

LATEST TRENDS on APPLIED MATHEMATICS, SIMULATION, MODELLING

(C) and non-linear registration (E), and the checkerboard images for the affine (D) and non-linear (F) registrations reveal a good quality alignment for both methods with a high degree of continuity of the structures.

H. Performance assessment

Fig. 4 Segmentation of spatial ICA components and white matter fiber tracts between the components. Shown are the thalamus IC26 (red) and IC3 (yellow) for one of the 100 subjects in the cohort. IC3 is a frontal lobe part of the default mode network. For assessment of structural connectivity between the functional regions IC26 and IC3, we extract the FA values from all voxels in the FA image that intersect those specific fibers (in blue).

G. Intra-subject, intra-modal registration of 3D MRI anatomy between wave-1 and wave-2 Healthy aging, and more so neurodegenerative disorders, are accompanied by changes in brain morphometry and volumetry. These changes, such as cortical thinning, subcortical shrinkage, and ventricular enlargements are not uniform across the brain and might differ substantially between individuals and point in lifetime when such changes are investigated. Therefore, nonlinear registration is particularly relevant in this case. Such successful nonlinear intra-modality registration is demonstrated in Fig. 5 using FSL’s flirt and fnirt.

III. DISCUSSION AND CONCLUSIONS The goal of the present work was to create a fully automated procedure for the image analysis in a large study of cognitive aging, where each subject dataset consisted of multimodal images (structural MRI, DTI, functional MRI). Before analyzing DTI data, it is recommended to perform an eddy-current correction by a set of affine registrations. Although the improvements might appear on a minor scale, the process is important not to compromise the quality of the derived measures (e.g. FA values). Beside the eddy current correction, we performed two intrasubject registrations: from the DTI domain and from the ICA domain into the anatomy. For these tasks, we used the FSL framework. FSL’s linear and nonlinear image registration (flirt and fnirt, respectively) performed well on MRI data when applied intra-modality. Nonlinear registration is highly desirable when accounting for small local deformations and to assess displacement fields inside and between anatomical structures. The latter could reveal the localization and progression of anatomical changes. However, FSL’s nonlinear

Fig. 5 Intra-modality affine and non-linear registration. The same subject was examined with structural MRI in wave-1 (A) and wave-2 (B). The 3years time separated image recordings were aligned using affine registration

ISSN: 1792-4332

Objective registration evaluation is an important issue. There are several options, for instance manually placed landmarks, variance measures on standard templates [16] or computation of image properties within the same region [19]. Here, we propose a slightly different approach for registration evaluation based on FreeSurfer segmentations. For our data, let A1 and A2 be the structural images and let S1 and S2 be the FreeSurfer segmentations from the wave-1 and wave-2, respectively (see Fig.5). Provided reliable segmentations S1 and S2, one can compute the percentage volume change of a specific region between those two time-points and use this as the “truth” for the volume change. Let T1,2 be the obtained nonlinear transformation from A1 to A2 (applied to A and giving E in Fig. 5 and apply T1,2 to S1. The obtained percentage change should then be equal to the “true” value. For one subject, we applied this approach to the hippocampus, which is expected to decrease in volume, and to the ventricles, which are expected to increase, and we obtained a percentage change of −4.2% and 1.86% for the respective registered anatomical regions, whereas the “truth” was −2.42% and 2.59%. As an internal control we applied the same approach to the affine registration between S1 and S2, and the values of −3.73% and 1.32% were obtained. Thus, we conclude that the major discrepancy between the affine registration and the “gold-standard” was due to interpolation effects, and since the nonlinear registration was within the same error range we conclude that it was successful.

267

ISBN: 978-960-474-210-3

LATEST TRENDS on APPLIED MATHEMATICS, SIMULATION, MODELLING

registration, fnirt, seems not to be tailored for inter-modality registration and does not necessarily produce acceptable results, even with an affine initialization. FSL’s nonlinear registration, fnirt, uses the sum of squared differences as cost functional (with an intensity modulation option), and is indeed doomed to work satisfactory only for images of the same modality. On the other hand, the affine registration flirt can use other discrepancy measures as cost functional, like mutual information. We also tested ITK’s (http://www.itk.org) nonlinear registration algorithms for the inter-modality registration, but we were not successful with parameter tuning to obtain fully satisfactory results. However, there are other options available for nonlinear inter-modality registration. As an alternative to intra-subject registration, it is possible to perform inter-subject registration into a standard space (i.e. MNI or Talairach) with existing anatomical and diffusion-weighted templates [16]-[19]. Instead of registration of DTI-derived images (i.e. the FA image) into the standard space one can perform nonlinear inter-modality tensor registration to a tensor-template [18], [21], [22]. However, there is a theoretical possibility of averaging subject specific information for nonlinear standard space registration since all subjects are registered to the same target. To account for this, one can perform intra-subject registration based on multichannel mutual information within each subject, which utilizes the relative independence of the DTI components [23], [24]. Presumably, this approach has reduced risk of such averaging. There are other considerations to take into account when opting for a linear/nonlinear registration method in such a large study: (i) that nonlinear transformations typically have more parameters to deal with and optimize; and (ii) that nonlinear transformations typically are more time consuming than affine ones. The latter is an important factor: unless nonlinear registration becomes computationally more efficient, it will be hard to advocate its use for large datasets. When opting for inter-modality affine transformations, we had again two choices: (i) perform a “header-blind” registration, or (ii) a registration based on spatial information in the NIFTI header. The “header-blind” registration performed normally well, but failed occasionally. The “header-based” registration, using header information about scanner coordinates, gave stable results for all subjects. Therefore, to ensure a robust, high-quality registration across the cohort, our choice was to perform an initial “header-based” registration, in order to obtain a good initial guess, followed by an affine registration for fine-tuning and improved accuracy. For ICAspace to anatomical-space registration, there was no such header information available, and the only option was an “uninformed” registration from the EPI template to the anatomy. Finally, in projects like the present one, with a large number of participants and several study waves, it is important to design automated analysis tools in contrast to manual subjectby-subject analysis. The advantages of automated tools are evident with respect to reproducibility, consistent errorgeneration, and easy and efficient adjustments of parameter settings. In this multidisciplinary study of “cognitive aging” we

ISSN: 1792-4332

268

have successfully implemented such integrated, automated workflow for the co-analysis of anatomical images, diffusion tensor images, and functional MRI time series, and in which well-performing and time-efficient registration procedures are a crucial component.

ACKNOWLEDGMENT The authors want to thank MedViz.uib.no for the interdisciplinary working environment, and the ”Cognitive aging” project (Prof. Astri J. Lundervold) for providing data. REFERENCES [1]

[2]

[3]

[4]

[5]

[6]

[7]

[8]

[9]

[10] [11]

[12]

[13]

[14]

[15]

[16]

Raichle, M.E., MacLeod, A.M., Snyder, A.Z., Powers, W.J., Gusnard, D.A., Shulman, G.L.: A default mode of brain function. Proc Natl Acad Sci U S A 98(2) (Jan 2001) 676–682. Bullmore, E., Sporns, O.: Complex brain networks: graph theoretical analysis of structural and functional systems. Nat Rev Neurosci 10(3) (Mar 2009) 186–198. Dale, A.M., Fischl, B., Sereno, M.I.: Cortical surface-based analysis. I. segmentation and surface reconstruction. Neuroimage 9(2) (Feb 1999) 179–194. Desikan, R.S., Ségonne, F., Fischl, B., Quinn, B.T., Dickerson, B.C., Blacker, D., Buckner, R.L., Dale, A.M., Maguire, R.P., Hyman, B.T., Albert, M.S., Killiany, R.J.: An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest. Neuroimage 31(3) (Jul 2006) 968–980. Ségonne, F., Dale, A.M., Busa, E., Glessner, M., Salat, D., Hahn, H.K., Fischl, B.: A hybrid approach to the skull stripping problem in MRI. Neuroimage 22(3) (Jul 2004) 1060–1075. Madden, D.J., Bennett, I.J., Song, A.W.: Cerebral white matter integrity and cognitive aging: contributions from diffusion tensor imaging. Neuropsychol Rev 19(4) (Dec 2009) 415–435. Sullivan, E.V., Adalsteinsson, E., Pfefferbaum, A.: Selective age-related degradation of anterior callosal fiber bundles quantified in vivo with fiber tracking. Cereb Cortex 16(7) (Jul 2006) 1030–1039. Damoiseaux, J.S., Beckmann, C.F., Arigita, E.J.S., Barkhof, F., Scheltens, P., Stam, C.J., Smith, S.M., Rombouts, S.A.R.B.: Reduced resting-state brain activity in the ”default network” in normal aging. Cereb Cortex 18(8) (Aug 2008) 1856–1864. Koch, W., Teipel, S., Mueller, S., Buerger, K., Bokde, A.L.W., Hampel, H., Coates, U., Reiser, M., Meindl, T.: Effects of aging on default mode network activity in resting state fMRI: Does the method of analysis matter? Neuroimage 51(1) (May 2010), 280-287. Yanovsky, I., Leow, A.D.: An information theory approach to nonlinear image registration. UCLA CAM Report 06-64 (2006). Xie, Z., Farin, G.E.: Image registration using hierarchical b-splines. IEEE Transactions on visualization and computer graphics 10(1) (2004) 85–94. Wu, M., Carmichael, O., Lopez-Garcia, P., Carter, C.S., Aizenstein, H.J.: Quantitative comparison of AIR, SPM and the fully deformable model for atlas-based segmentation of functional and structural MR images. Human Brain Mapping 27(9) (2006) 747–754. Collins, D.L., Neelin, P., Peters, T.M., Evans, A.C.: Automatic 3D intersubject registration of MR volumetric data in standardized Talairach space. Journal of Computer Assisted Tomography 18(2) (1994) 192–205. Woods, R.P., Grafton, S.T., Watson, J.D., Sicotte, N.L., Mazziotta, J.C.: Automated image registration: II. intersubject validation of linear and nonlinear models. Journal of Computer Assisted Tomography 22(1) (1998) 53–6515. Slomka, P.J., Dey, D., Przetak, C., Aladl, U.E., Baum, R.P.: Automated 3-dimensional registration of stand-alone 18F-FDG whole-body PET with CT. The Journal of Nuclear Medicine 44(7) (2003) 1156–1167. Geng, X., Christensen, G.E., Gu, H., Ross, T.J., Yang, Y.: Implicit reference-based group-wise image registration and its application to structural and functional MRI. Neuroimage 47(4) (Oct 2009) 1341– 1351.

ISBN: 978-960-474-210-3

LATEST TRENDS on APPLIED MATHEMATICS, SIMULATION, MODELLING

[17] Peng, H., Orlichenko, A., Dawe, R.J., Agam, G., Zhang, S., Arfanakis, K.: Development of a human brain diffusion tensor template. Neuroimage 46(4) (Jul 2009) 967–980. [18] Chiang, M.C., Leow, A.D., Klunder, A.D., Dutton, R.A., Barysheva, M., Rose, S.E., McMahon, K.L., de Zubicaray, G.I., Toga, A.W., Thompson, P.M.: Fluid registration of diffusion tensor images using information theory. IEEE Trans Med Imaging 27(4) (Apr 2008) 442– 456. [19] Mller, H.P., Unrath, A., Ludolph, A.C., Kassubek, J.: Preservation of diffusion tensor properties during spatial normalization by use of tensor imaging and fibre tracking on a normal brain database. Phys Med Biol 52(6) (Mar 2007) N99–109. [20] Pfefferbaum, A., Sullivan, E.V., Hedehus, M., Lim, K.O., Adalsteinsson, E., Moseley, M.: Age-related decline in brain white matter anisotropy measured with spatially corrected echo-planar diffusion tensor imaging. Magn Reson Med 44(2) (Aug 2000) 259–26. [21] Irfanoglu, M.O., Machiraju, R., Sammet, S., Pierpaoli, C., Knopp, M.V.: Automatic deformable diffusion tensor registration for fiber population analysis. Med Image Comput Comput Assist Interv 11(Pt 2) (2008) 1014–1022. [22] Barmpoutis, A., Vemuri, B.C., Forder, J.R.: Registration of high angular resolution diffusion MRI images using 4th order tensors. Med Image Comput Comput Assist Interv 10(Pt 1) (2007) 908–915. [23] Studholme, C.: Dense feature deformation morphometry: Incorporating DTI data into conventional MRI morphometry. Med Image Anal 12(6) (Dec 2008) 742–751. [24] Studholme, C.: Incorporating DTI data as a constraint in deformation tensor morphometry between T1 MR images. Inf Process Med Imaging 20 (2007) 223–232..

ISSN: 1792-4332

269

ISBN: 978-960-474-210-3