{"responseHeader":{"status":0,"QTime":7,"params":{"q":"{!q.op=AND}id:\"712693\"","hl":"true","hl.simple.post":"","hl.fragsize":"5000","fq":"!embargo_tdt:[NOW TO *]","hl.fl":"ocr_t","hl.method":"unified","wt":"json","hl.simple.pre":""}},"response":{"numFound":1,"start":0,"docs":[{"ark_t":"ark:/87278/s6bw0rrj","setname_s":"ir_uspace","restricted_i":0,"department_t":"","format_medium_t":"application/pdf","creator_t":"Gerig, Guido","identifier_t":"uspace,18961","date_t":"2014-01-01","bibliographic_citation_t":"Fishbaugh, J., Prestawa, M., Gerig, G., & Durrieman, S. (2014). Geodesic regression of image and shape data for improved modeling of 4D trajectories. IEEE International Symposium on Biomedical Imaging (ISBI), 385-8.","mass_i":1515011812,"publisher_t":"Institute of Electrical and Electronics Engineers (IEEE)","description_t":"A variety of regression schemes have been proposed on images or shapes, although available methods do not handle them jointly. In this paper, we present a framework for joint image and shape regression which incorporates images as well as anatomical shape information in a consistent manner. Evolution is described by a generative model that is the analog of linear regression, which is fully characterized by baseline images and shapes (intercept) and initial momenta vectors (slope). Further, our framework adopts a control point parameterization of deformations, where the dimensionality of the deformation is determined by the complexity of anatomical changes in time rather than the sampling of the image and/or the geometric data. We derive a gradient descent algorithm which simultaneously estimates baseline images and shapes, location of control points, and momenta. Experiments on real medical data demonstrate that our framework effectively combines image and shape information, resulting in improved modeling of 4D (3D space + time) trajectories.","first_page_t":"385","rights_management_t":"(c) 2014 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.","title_t":"Geodesic regression of image and shape data for improved modeling of 4D trajectories","id":712693,"publication_type_t":"pre-print","parent_i":0,"type_t":"Text","thumb_s":"/8e/20/8e20038b873a81d6ca31f452622f87d75bcc8090.jpg","last_page_t":"388","oldid_t":"uspace 10706","metadata_cataloger_t":"CLR","format_t":"application/pdf","modified_tdt":"2014-10-17T00:00:00Z","school_or_college_t":"","language_t":"eng","file_s":"/45/71/45716b039332913438de66d356b4e345f5e6df15.pdf","format_extent_t":"1,223,409 bytes","other_author_t":"Fishbaugh, James; Prestawa, Marcel; Durrieman, Stanley","created_tdt":"2014-10-17T00:00:00Z","_version_":1642982791787315200,"ocr_t":"GEODESIC REGRESSION OF IMAGE AND SHAPE DATA FOR IMPROVED MODELING OF 4D TRAJECTORIES James Fishbaugh1, Marcel Prastawa1, Guido Gerig1*, Stanley Durrleman2 1Scientific Computing and Imaging Institute, University of Utah 2INRIA/ICM, Piti´e Salpˆetri`ere Hospital, Paris, France ABSTRACT A variety of regression schemes have been proposed on im-ages or shapes, although available methods do not handle them jointly. In this paper, we present a framework for joint image and shape regression which incorporates images as well as anatomical shape information in a consistent manner. Evolution is described by a generative model that is the analog of linear regression, which is fully characterized by baseline images and shapes (intercept) and initial momenta vectors (slope). Further, our framework adopts a control point pa-rameterization of deformations, where the dimensionality of the deformation is determined by the complexity of anatom-ical changes in time rather than the sampling of the image and/or the geometric data. We derive a gradient descent al-gorithm which simultaneously estimates baseline images and shapes, location of control points, and momenta. Experi-ments on real medical data demonstrate that our framework effectively combines image and shape information, resulting in improved modeling of 4D (3D space + time) trajectories. 1. INTRODUCTION Analysis of longitudinal data incorporating both spatial and temporal information is essential for various clinical tasks such as predicting patient outcome and measuring efficacy of different therapeutic strategies. A crucial tool for longi-tudinal analysis is regression of observed data, which enables interpolation to generate continuous evolution models as well as extrapolation to predict future observations. Regression models are also necessary for conducting population studies comparing the change trajectories of different subjects. In medical imaging, it is important to consider image data in anatomical context, which motivates regression on image and shape data in different combinations (a multi-object com-plex). A variety of regression schemes have been proposed on images or shapes, although available methods do not handle them jointly. For example, the extension of kernel regression *Supported by grants: RO1 HD055741 (ACE, project IBIS), U54 EB005149 (NA-MIC), and U01 NS082086 (HD) and the Utah Science Tech-nology and Research (USTAR) initiative at the University of Utah. for image data [1] or piecewise linear regression for time se-ries of images [2] and shapes [3]. Combining intensity and geometric information has been explored for registration [4]. To conduct statistical analysis on 4D (3D space + time) data, it is particularly useful to consider compact generative regression models which have low number of parameters sit-uated at only one chosen time point. Geodesic regression is such a model and is fully characterized by baseline images and shapes (the intercept) and the tangent vector defining the geodesic at the baseline objects (the slope). Geodesic regres-sion frameworks for images [5, 6] and for shapes [7] have been proposed using the LDDMM setting. However, no no-tion of how to combine images and shape data is provided. We propose a novel geodesic regression framework that leverages image and shape data together to estimate a sin-gle deformation of the ambient space. We use the currents representation for geometric data that allows flexible repre-sentation of a wide variety of shape objects such as point sets, curves, or surface meshes, without the need for point correspondence between shapes. Compared to image regres-sion alone, shape data provides anatomical information that constrains the regression, especially in cases where images have low contrast, by placing larger weights on regions with anatomical importance. Compared to shape regression alone, image information provides data in areas where segmenta-tions are not available, as well as providing context to re-gions surrounding anatomical objects. Our framework uses the control-point parameterization of geodesic flows intro-duced in [8], which makes the parameterization of the defor-mation independent from the data. This allows us to keep a reasonable dimension of the parameterization, which is deter-mined by the complexity of anatomical changes in time, and not the sampling of the data. We therefore combine image and shape data without introducing a complexity overhead. 2. METHODOLOGY We performregression on observed time-series data of images Ii and shapes Xi, each acquired at time-point ti. Shape data may consist of a mix of point sets, curves, or surface meshes where all vertices are concatenated into one vector Xi. 978-1-4673-1961-4/14/$31.00 ©2014 IEEE 385 We use the control point formulation of [8] to generate geodesic flows of diffeomorphisms. Let S0 = {c0,k, 0,k} be a set of momentum vectors 0,k attached to control points c0,k distributed in the image domain. Geodesic flows are com-puted by evolving control points and momenta by integrating the following Hamiltonian equations over the time interval of interest: ˙ ck(t) = N Xi=1 K(ck(t), ci(t))i(t) ˙ k(t) = − N Xi=1 k(t)Ti(t)∇1K(ck(t), ci(t)) (1) with initial conditions ck(0) = c0,k and k(0) = 0,k (as-suming starting time-point to be 0), and K is a Gaussian ker-nel with variance 2V which controls the spatial scale of de-formation. For simplicity, we write these equations as ˙S (t) = F(S(t)) with S(0) = S0. The convolution of the momenta defines the following time-varying velocity field: v(t, x) = PN i=1 K(x, ci(t))i(t) for any point x in the domain. The velocity is used to deform the domain: a particle at point x at time 0 moves to (t, x) at later time t, where (t, x) fol-lows the integral curve of @(t,x) @t = v(t, (t, x)) starting with (0, x) = x. In this formulation, the velocity of the particle is given by the field v(t, .) at its current location. It has been shown in [9] that for all t, (t, .) is a 3D diffeomorphism. Following this model, the vertices of a given baseline shape complex concatenated into a vector X0 move at time t to X(t) = (t,X0), which satisfies the ordinary differential equation (ODE): ˙X (t) = v(t,X(t)) with X(0) = X0. To make explicit the dependency of the equation of motion on S(t), we write it as: ˙X (t) = G(X(t), S(t)). A given baseline image I0 is also deformed by the flow Fig. 1. Conceptual overview of geodesic regression on multi-object complexes containing both image and shape data. The framework estimates parameters at t = 0 which consist of the baseline image I0 and shape X0 along with the deformation model parameterized by control points c0 and initial momenta 0 such that overall distance between the deformed objects and the observations are minimal. of diffeomorphisms and its trajectory is given as I(t) = I0 ◦ (t, .)−1. The inverse flow satisfies the equation @(t,.)−1 @t = −d(t, .)−1v(t, .). For the sake of simplicity, we denote Y (t, .) = (t, .)−1, a L2 function that maps the point x to its position at time t under the inverse flow −1(t, x). This maps satisfies ˙Y (t, .) = −dY (t, .)v(t, .) = H(Y (t, .), S(t)), where we make explicit the dependency on S(t). At time t, the intensity of the warped baseline image at voxel position x is given by I(t, x) = I0(Y (t, x)) using 3D interpolation. A conceptual overview of our framework is shown in Fig. 1 where regression is performed by minimizing the overall distance between the observations and the deformed baseline objects (shapes and/or images). Let d(X(ti),Xi) be a metric between the deformed baseline shape complex X0 at time ti and the data shape complex Xi. This metric may be a weighted sum over each component of the shape complex of the currents metric between sets of curves or sur-face meshes. This term essentially depends on X(ti) and is denoted A(X(ti)). Similarly, we have a metric d(I(ti), Ii) denoted as B(Y (ti, .)) that is the sum of squared differences between the deformed baseline image I0 ◦ Y (ti, .) and the observed image Ii. The geodesic regression problem amounts to finding the deformation parameters S0 and baseline anatomical configu-ration (I0,X0) such that the following criterion is minimized: E(S0, I0,X0) =Xt i Iti A(X(ti)) + Sti B(Y (ti, .)) + L(S0) (2) subject to ˙S (t) = F(S(t)) S(0) = S0 ˙X (t) = G(X(t), S(t)) X(0) = X0 ˙Y (t, .) = H(Y (t, .), S(t)) Y (0, .) = Id (3) where the regularizer L(S0) = PN i,j=1 T 0,iK(c0,i, c0,j)0,j is the squared norm of initial velocity and weights on image and shape matching Iti and Sti . As shown in the supplementalmaterial (www.cs.utah. edu/˜jfishbau/docs/isbi2014_eqns.pdf), the gradient is computed by integrating 3 linear ODEs with source terms from final time-point Tf back to time-point 0: ∇S0E = (0) + ∇S0L ∇X0E = (0) 1 2 ∇I0E =Xt j SplatY (tj ,.)(I0 ◦ Y (tj , .) − Ii) with ˙(t) = −@1G(t)T (t) −Xt i ∇X(ti)A(t − ti) ˙(t) = −@1H(t)†(t) −Xt i ∇Y (ti,.)B (t − ti) ˙(t) = −@2G(t)T (t) − @2H(t)†(t) − dS(t)FT (t) 386 with final conditions (Tf ) = (Tf ) = (Tf ) = 0. The vector is same size asX0, which brings back to time t = 0 the gradients of the data matching terms, and is used to update the position of the vertices of the baseline shape com-plex. Similarly is of the same size as Y (0, .) (an image of vectors in practice) which integrates the successive gradients of the image matching terms that acts as jumps in the differ-ential equation. Finally, is a variable of the same size as S0 which is used at time t = 0 to update the deformation param-eters (the position of the control points and their momentum vectors). The gradient with respect to the baseline image in-volves the splatting of the current residual images at positions Y (ti, .) as done in [8]. 3. RESULTS AND DISCUSSION Pediatric Brain Development: We explore the impact of joint image and shape regression in modeling pediatric brain development. The data consists of T1W images of the same healthy child observed at 6, 12, and 25 months of age. Re-gression on images alone is difficult in this case due to the very low contrast in the 6 month old image. Despite the low contrast, tissue segmentations can still be reliably and consis-tently estimated [10]. We estimate a geodesic model using only T1W images and a model jointly on images and white matter surfaces to emphasize the development of the tissue interface. We initialize 120 control points on a regular grid with the deformation kernel V = 20 mm. Finally, due to limited contrast at 6 months, we estimate the baseline at 25 months and follow the evolution backwards in time. The results of geodesic regression are shown for several snapshots in time in Fig. 2. The model estimated using only images mostly captures the scale change, but does not cap- Fig. 2. Images and deformations estimated by geodesic re-gression using images alone (top) and jointly on images and white matter surfaces (bottom). Regression jointly on image and shape results in a more realistic evolution which captures detailed changes in brain tissue in addition to the increase in brain size. In both cases, geodesic regression was estimated backwards in time. 0 5 10 15 20 25 3200 3300 3400 3500 3600 Model estimated from images alone Caudate Volume (mm3) Time from baseline (months) Left Right 0 5 10 15 20 25 3200 3300 3400 3500 3600 Model estimated jointly on image and shapes Caudate Volume (mm3) Time from baseline (months) Left Right Fig. 3. Caudate volume extracted continuously after regres-sion compared to observed caudate volumes (circles and x's). Volume is measured continuously from the modeled shape trajectories, not fitted to discrete volume measurements. The model estimated on images alone fails to capture the volume loss. Evolution of caudates for the image only model is not estimated, but instead we shoot the baseline caudate shapes along the geodesic estimated from images alone. Note: mea-surements extracted continuously from non-linearly deform-ing shapes can produce either linear or non-linear trends with no prior assumption of linearity. ture much deformation in the interior of the brain. The model estimated jointly on image and shape captures more detailed development as white matter stretches and expands. Neurodegeneration in Huntington's Disease: Next, we in-vestigate the application of joint image and shape regression to Huntington's disease (HD) where accurate 4D models are needed to measure the effectiveness of therapies or drug treat-ments. In HD, degeneration of the caudate has been shown to be significant [11]. Here we explore T1W image data from a single patient diagnosed with HD scanned at 58, 59, and 60 years of age. Sub-cortical structures are segmented, manually verified, and cleaned. Models are estimated using only T1W images as well as T1Wimages plus caudate surfaces. Control points are initialized on a regular grid with 10 mm spacing with kernel V = 10 mm. The trajectory of caudate volume extracted after regres-sion is shown in Fig. 3. The model estimated from images alone fails to capture the volume loss observed in both cau-dates, and rather, shows an increase in right caudate volume. By incorporating caudate shape data in model estimation, we 387 Fig. 4. Top) Evolution estimated on images alone. Evolution of caudates are not estimated, but instead we shoot the baseline caudate shapes along the estimated geodesic. Bottom) Evolution estimated jointly using images and caudate shapes. Regression on images alone results in a slight expansion of ventricles, but does not capture the shrinking of caudates. Our method is able to capture both the expansion of ventricles and the shrinking of caudates. are able to capture the shrinking of the caudates. The corre-sponding expansion of the ventricles is also captured, shown in Fig 4, due to the inclusion of imaging data. By incorporat-ing shape and image information jointly, we are able to model both the expansion of the ventricles and the degeneration of the caudates. Accurate models of change are essential when extrapolating beyond the observation time interval, which can provide insight into disease progression. Conclusions: We presented a novel geodesic regression framework that jointly considers image and shape information in the LDDMM framework, where dense diffeomorphisms are built using a control point formulation. This formulation decouples deformation parameters from input object parame-ters (e.g., voxels, surface points) providing greater flexibility and consistency in mapping different object types across time. Our regression model seamlessly handles images and multi-object complexes consisting of points, curves, and/or surfaces in different combinations. Experiments show that our frame-work effectively combines image and shape information to estimate a single deformation of the ambient space, resulting in improved modeling of 4D trajectories. 4. REFERENCES [1] B.C. Davis, P.T. Fletcher, E. Bullitt, and S. Joshi, \"Pop-ulation shape regression from random design data,\" in ICCV. 2007, pp. 1-7, IEEE. [2] A.R. Khan and M.F. Beg, \"Representation of time-varying shapes in the large deformation diffeomorphic framework,\" in ISBI. 2008, pp. 1521-1524, IEEE. [3] S. Durrleman, X. Pennec, A. Trouv´e, J. Braga, G. Gerig, and N. Ayache, \"Toward a comprehensive framework for the spatiotemporal statistical analysis of longitudinal shape data,\" IJCV, pp. 1-38, 2012. [4] P. Cachier, E. Bardinet, D. Dormont, X. Pennec, and N. Ayache, \"Iconic feature based nonrigid registration: the pasha algorithm.,\" Computer Vision and Image Un-derstanding, vol. 89, no. 2-3, pp. 272-298, 2003. [5] M. Niethammer, Y. Huang, and F.X. Vialard, \"Geodesic regression for image time-series,\" in MICCAI, 2011, vol. 6892 of LNCS, pp. 655-662. [6] N.P. Singh, J. Hinkle, S. Joshi, and P.T. Fletcher, \"A vector momenta formulation of diffeomorphisms for im-proved geodesic regression and atlas construction,\" in ISBI, 2013, pp. 1219-1222. [7] J. Fishbaugh, M. Prastawa, G. Gerig, and S. Durrle-man, \"Geodesic shape regression in the framework of currents,\" in IPMI, 2013, vol. 23, pp. 718-729. [8] S. Durrleman,M. Prastawa, G. Gerig, and S. Joshi, \"Op-timal data-driven sparse parameterization of diffeomor-phisms for population analysis,\" in IPMI, 2011, vol. 6801 of LNCS, pp. 123-134. [9] S. Joshi and M. Miller, \"Landmark matching via large deformation diffeomorphisms,\" IEEE Trans. Image Pro-cessing, vol. 9, no. 8, pp. 1357-1370, 2000. [10] A. Vardhan, M. Prastawa, S. Gouttard, J. Piven, and G. Gerig, \"Quantifying regional growth patterns through longitudinal analysis of distances between mul-timodal MR intensity distributions,\" in ISBI, 2012, pp. 1156-1159. [11] EH. Aylward, AM. Codori, A. Rosenblatt, M. Sherr, J. Brandt, OC. Stine, PE. Barta, GD. Pearlson, and CA. Ross, \"Rate of caudate atrophy in presymptomatic and symptomatic stages of huntington's disease,\" Movement Disorders, vol. 15, no. 3, pp. 552-560, 2000. 388"}]},"highlighting":{"712693":{"ocr_t":[]}}}