fully automated whole-body registration in mice …
TRANSCRIPT
FULLY AUTOMATED WHOLE-BODY REGISTRATION IN MICE USING AN
ARTICULATED SKELETON ATLASFULLY AUTOMATED WHOLE-BODY REGISTRATION
IN MICE USING AN ARTICULATED SKELETON ATLAS
M. Baiker1,2, J. Milles1, A. M. Vossepoel2, I. Que3, E. L. Kaijzel3, C. W. G. M. Löwik3,
J. H. C. Reiber1, J. Dijkstra1 and B. P. F. Lelieveldt1
1Division of Image Processing Department of Radiology
Leiden University Medical Center, The Netherlands
2Quantitative Imaging Group Delft University of Technology
The Netherlands
ABSTRACT
In this paper, we propose a fully automated articulated registration approach for whole-body 3D data of mice. The method is based on a hierarchical anatomical model of the skeletal system where we specified position and degrees of freedom for each joint. Model fitting is performed by traversing a hierarchical part-tree, which enables a coarse- to-fine registration from the inner articulations outwards. The method was tested on 12 Micro-CT volumes, giving accurate alignment of the skeletal structures in all cases.
Index Terms— Articulated registration, Micro-CT, atlas-based matching
1. INTRODUCTION Micro-CT enables non-invasive monitoring of disease development and treatment response over time in the whole animal body, including the extremities and the head. However, inter-scan variations in position of the limbs due to non-uniform animal positioning complicate accurate comparison of disease status between time points. They also hamper morphometric intra- and inter-subject comparison. To compensate for such variations between individuals and time-instances, a body-to-body registration step is required. One way to handle this has been introduced by Kovacevic et al. [1]. Based on the “part-of” concept they use a hierarchical approach, separating first the main organ compounds and refining that division as the registration progresses, down to single bones and organs. However, this approach uses the hierarchical framework only for initialization of registrations on lower levels. Moreover, the method can only capture minor postural differences and was tested on simulated data. To handle arbitrary shape variations, relations between connected objects have to be taken into account not only for initialization but also during the registration step itself.
Earlier works on articulated registration focused on subparts of the body like the legs of a mouse, based on the leg bones [2] and a human head, based on a part of the spine [3]. Soft tissue parts were included using a continuous deformation. In this paper, we propose a fully automated, articulated registration method for aligning the entire skeleton of mice, whose postures can differ significantly. We chose the skeleton because it forms the rigid frame of a body and it is the main determinant of whole-body shape. As a consequence, aligning the skeletons of two subjects captures most of the difference in posture and allows subsequent initialization of the registration of the rest of the body (e.g. the organs). Moreover, the skeleton can be robustly and automatically detected in Micro-CT data, yielding a reliable feature for registration. The proposed approach is based on an anatomical mouse atlas developed by Segars et al. [4], which contains the mouse skeleton as well as major organs. Departing from this atlas, we define a hierarchical anatomical model of the skeletal system. The contributions of this paper are twofold:
We present an improved digital mouse atlas where the skeletal system is divided into intuitive bone compounds, with joints and realistic articulation Degrees of Freedom (DoF).
We combine the “part-of” concept and articulated registration to fully automatically register the aforementioned atlas to the skeleton of a given mouse that is extracted from a Micro-CT dataset.
2. METHODOLOGY
2.1. Hierarchical anatomical model The hierarchical anatomical tree used for this work is shown in Figure 1. The strategy for capturing the bone structure is first to coarsely align the atlas with the real skeleton and then to apply an articulated registration scheme traversing the hierarchical tree. This way, the lower tree levels are
7281424406722/07/$20.00 ©2007 IEEE ISBI 2007
initialized and constrained by the higher level registration transformations. Therefore, the highest hierarchical level is the entire mouse skeleton itself. This top level can be divided into subparts that consist either of single bones or bone compounds. The skull is placed on a higher level than the other elements of the skeleton since the matching procedures for all the other parts are initialized by the result of the skull registration. Subsequently the rear part consisting of pelvis, upper and lower hind limbs, and the paws can be initialized by means of a rigid connection between spine and pelvis. The registration of the front limbs that are divided into upper and lower limbs and the paws is directly initialized by the skull. Further distinctions like refinement of the paws are not relevant for the goal of capturing the animal posture. Assuming that the spine and the sternum sufficiently represent the rib cage, further distinction into single ribs can equally be avoided.
Entire body (skeleton)
Pelvis
Paw (left and right)
Paw (left and right)
Front limbsHind limbs
Figure 1: Hierarchical anatomical tree for the entire skeletal system. The connections depict immediate relations between two elements of element groups i.e. a connected part on a lower level is initialized by the registration result on a higher level
2.2. Modifications of the mouse atlas The model used for the skeleton registration is the MRI- based 4D digital mouse atlas presented in [4]. However, the skeleton that is included in the atlas does not contain a distinction between single bones and joints. Therefore, the bones were segmented manually from the skeleton using Amira V3.1 [5], guided by an anatomical text book [6] and a high resolution CT scan of a real mouse (Figure 2). Second, the position and the DoFs were specified for each joint. We distinguished three types of joints: ball and hinge joints and the shoulder complex (both shoulders combined). Table 1 displays the DoFs for the ball and hinge joints.
Figure 2: The mouse skeleton as included in the original atlas [4] (top) and after segmentation of single bones and adding joints (bottom) Due to the many DoFs in the shoulder complex, we introduce an additional motion constraint for the shoulder by first allowing only a coupled, symmetric displacement of both front upper limbs, with a varying distance between the shoulders and a rotation towards and away from each other. Subsequently, the individual shoulders are decoupled and treated as ball joints, with 9 DoFs in the registration. For each joint, these pre-specified DoFs are assigned to the corresponding node in the hierarchical anatomical tree: these serve as kinematic constraints during the tree matching.
Joint types Modeled joint DoFs of the articulated bone
Ball joint
Hinge joint
Elbow Knee
3 translations 1 rotation 3 scalings
Table 1: Joint types in the atlas skeleton and the DoFs for the registration of the distal articulated bone (joint pictograms from [7])
2.3. Articulated registration To fit the different articulations, we convert both the CT data and the atlas into a surface representation. Subsequently, we apply the iterative closest point (ICP) algorithm [8], which minimizes the Euclidean distance between two point sets. The articulated registration is performed by traversing the hierarchical anatomical tree in a top-down manner, optimizing the DoFs specified in each model node. After convergence for a node level, the error criterion is further minimized for the lower node level, yielding a gradually decreasing error function. Depending on the joint type, these DoFs differ per node. However, all registration steps include translations, a varying number of rotations and anisotropic scaling to capture possible differences in size between individuals (see Table 1). The error criterion is minimized with respect to the current node parameters using Levenberg-Marquardt minimization.
x
y
z
729
An exception to this scheme is the spine which is not determined by registration but by binning the bone point set along the longitudinal axis and applying three dimensional region growing, starting from the head-spine connection. The amount of points in each bin increases significantly when the spine-pelvis connection is reached. This initializes the pelvis registration. 2.4. Coarse whole-body alignment To initialize the articulated registration, the mouse model needs to be coarsely aligned to the skeleton segmented from CT i.e. global DoFs have to be removed. Figure 2 shows that the first principal axis of the bone voxels defines the longitudinal body axis (further referred to as z-axis). By binning the bone voxels along the z-axis and computing the Center of Gravity (CoG) for each bin, we derive a 3D curve that enables coarse alignment in the following manner. First, a possible rotation around the longitudinal axis is resolved. Due to the relatively larger asymmetry of the skeleton in the vertical plane (yz-plane in Figure 2) compared to the horizontal plane (assuming that the limbs are placed somewhat symmetrical to the vertical plane), projecting the CoG curve on the xy-plane yields a 2D point set whose first eigenvalue is significantly larger than its second eigenvalue. Thus, the first eigenvector indicates the vertical direction. Second, prone or supine subject position can be derived from the CoG curve between the rib cage and the pelvis. Third, the front and hind part of the subject can be distinguished by assessing the amount of bone along the z- axis where the front shows a peak in bone density due to the skull. Finally, a possible translation along the z-axis is
resolved using a significant maximum of the CoG curve in y around the head-spine connection. An initial scale factor is derived from the relative bone volume of the CT skeleton compared to the atlas bone volume. The initial alignment parameters are stored in the top node of the hierarchical part tree.
3. EXPERIMENTAL RESULTS
To test the registration framework, 3D data volumes were acquired of 12 mice with different postures with a Skyscan 1178 Micro-CT scanner. The original data resolution was 80x80x80 m3. The data was sub-sampled with a factor 4, smoothed and the skeleton was segmented through isodata thresholding. While the coarse structure of all bones that are included in the registration process is retained, all the ribs have been removed by the preprocessing steps to ensure robust matching of the sternum. Since the shoulder blades are very thin, they are not properly represented after the preprocessing steps and are therefore not considered. The resulting skeleton representation is subsequently used for the coarse alignment step as described in the previous section. Two representative registration results on mice with very different postures are shown in Figure 3. Isosurfaces of the segmented skeletons are shown in grey and the phantom skeleton surface in red. The upper and the middle row show the skeletons before and after coarse registration respectively. The last row shows the results after the articulated registration step. The spine is represented by the CoGs of the bins that were used for the 3D region growing.
Figure 3: Two examples (left and right column) of the registration between the model (red) and segmented micro-CT data (grey) before registration (top row), after the coarse alignment step (middle row) and after the articulated registration (bottom row)
730
00 Coarsely aligned skeleton 01 Skull 02 Right part of the pelvis 03 Right hind upper limb 04 Right hind lower limb 05 Right hind paw 06 Left part of the pelvis 07 Left hind upper limb 08 Left hind lower limb 09 Left hind paw 10 Breast bone 11 Right front upper limb 12 Right front lower limb 13 Right front paw 14 Left front upper limb 15 Left front lower limb 16 Left front paw
Figure 4: Decrease of the error criterion (minimum distance from a phantom surface node to a CT skeleton surface node) calculated including all phantom surface nodes (approx. 3000 nodes and approx. 15000 nodes for the CT skeleton surface) for all 12 datasets as the hierarchical tree is traversed (left) and mean and standard deviation of the average error for specific bones from all 12 datasets before and after registration (right)
All 12 cases yielded a similar correct registration between the CT-segmented skeleton and the atlas. The decrease in the registration error for all 12 datasets is plotted in Figure 4 (left). The registration error including all phantom surface nodes decreases from an average of 2.93 ±0.63mm to 0.44 ±0.04mm. Figure 4 (right) shows the averaged bone specific registration result for all 12 datasets. One can clearly see that the lower a bone is located in the hierarchy, the larger the error is before the registration. After registration, there is no relation anymore between error and hierarchical level. The entire registration procedure was implemented in Matlab 7.1 and took less than 2 seconds for the coarse registration and 2-3 minutes for the articulated registration on a standard desktop PC.
4. DISCUSSION AND CONCLUSION We presented a fully automated method for atlas-based segmentation of an entire mouse skeleton, and showed experimental results on real data. The chosen ICP algorithm, applied to the surface representations of CT data and the atlas, proved to lead to a registration result with an average surface node distance of 0.44 ±0.04mm, including all datasets. While in general the registration result is very good, there are some areas where parts of the atlas elements deviate from the real data. This may be caused by non-linear shape differences between the atlas and the data, which can be removed by applying a subsequent non-rigid registration. Also, implementation of a second iteration step that takes the first one as an initialization and/or using data at higher resolution may improve registration accuracy and allows including thinner bones like the shoulder blades or the ribs as well. Though computation time is generally acceptable, it could be improved by defining a volume of interest in the CT data prior to the articulated registration step, after an atlas element is initialized. This can also be realized by removing the model parts that have been already registered from the computations. Also, the “part-of”
concept allows parallel computing and therefore the calculation speed could be increased as well. Based on this work, we will focus on the integration of the atlas soft tissue elements such as the lungs, heart and intestinal organs that can be distinguished in CT data in the registration procedure. The skeleton based registration, combined with the skin surface then serves to constrain a non-rigid registration for the soft tissues, yielding a whole-body segmentation. Also, we expect that this method generalizes well towards other rodents, provided that an anatomical atlas is available.
5. ACKNOWLEDGEMENTS The authors gratefully acknowledge Henk Rozemuller and Maj Petersen for preparing animals for CT scanning, Dr Paul Segars for providing us with the mouse atlas and Elke van de Casteele from Skyscan for providing the Skyscan Micro-CT scanner used in this research.
6. REFERENCES
[1] N. Kovacevic, G. Hamarneh and M. Henkelman, Proc. MICCAI 2003, LNCS, vol 2879, pp 870-877
[2] X. Papademetris, D.P. Dione, L.W. Dobrucki, L.H. Staib and A.J. Sinusas, Proc. MICCAI 2005, LNCS, vol 3750, pp 919-926
[3] A. du Bois d’Aische, M. De Craene, B. Macq and S.K. Warfield, Proc. IEEE Conference on Image Processing 2005, vol 1, pp 21-24
[4] W.P. Segars, B.M.W. Tsui, E.C. Frey, G.A. Johnson, and S.S. Berr, “Development of a 4D digital mouse phantom for molecular imaging research”, Molecular Imaging and Biology 2004, vol 6(3), pp 149-159
[5] http://www.mc.com/tgs [6] M.J. Cook, “Anatomy of the Laboratory Mouse”,
ISBN 0121869504 [7] http://preventdisease.com/fitness/fundament/articles/ty
pes_of_joints.html (01/12/2006) [8] P.J. Besl and N.D. McKay, “A method for registration
of 3D shapes”, IEEE Transactions on Pattern Analysis and Machine Intelligence 1992, vol 14(2), pp 871-881
731
M. Baiker1,2, J. Milles1, A. M. Vossepoel2, I. Que3, E. L. Kaijzel3, C. W. G. M. Löwik3,
J. H. C. Reiber1, J. Dijkstra1 and B. P. F. Lelieveldt1
1Division of Image Processing Department of Radiology
Leiden University Medical Center, The Netherlands
2Quantitative Imaging Group Delft University of Technology
The Netherlands
ABSTRACT
In this paper, we propose a fully automated articulated registration approach for whole-body 3D data of mice. The method is based on a hierarchical anatomical model of the skeletal system where we specified position and degrees of freedom for each joint. Model fitting is performed by traversing a hierarchical part-tree, which enables a coarse- to-fine registration from the inner articulations outwards. The method was tested on 12 Micro-CT volumes, giving accurate alignment of the skeletal structures in all cases.
Index Terms— Articulated registration, Micro-CT, atlas-based matching
1. INTRODUCTION Micro-CT enables non-invasive monitoring of disease development and treatment response over time in the whole animal body, including the extremities and the head. However, inter-scan variations in position of the limbs due to non-uniform animal positioning complicate accurate comparison of disease status between time points. They also hamper morphometric intra- and inter-subject comparison. To compensate for such variations between individuals and time-instances, a body-to-body registration step is required. One way to handle this has been introduced by Kovacevic et al. [1]. Based on the “part-of” concept they use a hierarchical approach, separating first the main organ compounds and refining that division as the registration progresses, down to single bones and organs. However, this approach uses the hierarchical framework only for initialization of registrations on lower levels. Moreover, the method can only capture minor postural differences and was tested on simulated data. To handle arbitrary shape variations, relations between connected objects have to be taken into account not only for initialization but also during the registration step itself.
Earlier works on articulated registration focused on subparts of the body like the legs of a mouse, based on the leg bones [2] and a human head, based on a part of the spine [3]. Soft tissue parts were included using a continuous deformation. In this paper, we propose a fully automated, articulated registration method for aligning the entire skeleton of mice, whose postures can differ significantly. We chose the skeleton because it forms the rigid frame of a body and it is the main determinant of whole-body shape. As a consequence, aligning the skeletons of two subjects captures most of the difference in posture and allows subsequent initialization of the registration of the rest of the body (e.g. the organs). Moreover, the skeleton can be robustly and automatically detected in Micro-CT data, yielding a reliable feature for registration. The proposed approach is based on an anatomical mouse atlas developed by Segars et al. [4], which contains the mouse skeleton as well as major organs. Departing from this atlas, we define a hierarchical anatomical model of the skeletal system. The contributions of this paper are twofold:
We present an improved digital mouse atlas where the skeletal system is divided into intuitive bone compounds, with joints and realistic articulation Degrees of Freedom (DoF).
We combine the “part-of” concept and articulated registration to fully automatically register the aforementioned atlas to the skeleton of a given mouse that is extracted from a Micro-CT dataset.
2. METHODOLOGY
2.1. Hierarchical anatomical model The hierarchical anatomical tree used for this work is shown in Figure 1. The strategy for capturing the bone structure is first to coarsely align the atlas with the real skeleton and then to apply an articulated registration scheme traversing the hierarchical tree. This way, the lower tree levels are
7281424406722/07/$20.00 ©2007 IEEE ISBI 2007
initialized and constrained by the higher level registration transformations. Therefore, the highest hierarchical level is the entire mouse skeleton itself. This top level can be divided into subparts that consist either of single bones or bone compounds. The skull is placed on a higher level than the other elements of the skeleton since the matching procedures for all the other parts are initialized by the result of the skull registration. Subsequently the rear part consisting of pelvis, upper and lower hind limbs, and the paws can be initialized by means of a rigid connection between spine and pelvis. The registration of the front limbs that are divided into upper and lower limbs and the paws is directly initialized by the skull. Further distinctions like refinement of the paws are not relevant for the goal of capturing the animal posture. Assuming that the spine and the sternum sufficiently represent the rib cage, further distinction into single ribs can equally be avoided.
Entire body (skeleton)
Pelvis
Paw (left and right)
Paw (left and right)
Front limbsHind limbs
Figure 1: Hierarchical anatomical tree for the entire skeletal system. The connections depict immediate relations between two elements of element groups i.e. a connected part on a lower level is initialized by the registration result on a higher level
2.2. Modifications of the mouse atlas The model used for the skeleton registration is the MRI- based 4D digital mouse atlas presented in [4]. However, the skeleton that is included in the atlas does not contain a distinction between single bones and joints. Therefore, the bones were segmented manually from the skeleton using Amira V3.1 [5], guided by an anatomical text book [6] and a high resolution CT scan of a real mouse (Figure 2). Second, the position and the DoFs were specified for each joint. We distinguished three types of joints: ball and hinge joints and the shoulder complex (both shoulders combined). Table 1 displays the DoFs for the ball and hinge joints.
Figure 2: The mouse skeleton as included in the original atlas [4] (top) and after segmentation of single bones and adding joints (bottom) Due to the many DoFs in the shoulder complex, we introduce an additional motion constraint for the shoulder by first allowing only a coupled, symmetric displacement of both front upper limbs, with a varying distance between the shoulders and a rotation towards and away from each other. Subsequently, the individual shoulders are decoupled and treated as ball joints, with 9 DoFs in the registration. For each joint, these pre-specified DoFs are assigned to the corresponding node in the hierarchical anatomical tree: these serve as kinematic constraints during the tree matching.
Joint types Modeled joint DoFs of the articulated bone
Ball joint
Hinge joint
Elbow Knee
3 translations 1 rotation 3 scalings
Table 1: Joint types in the atlas skeleton and the DoFs for the registration of the distal articulated bone (joint pictograms from [7])
2.3. Articulated registration To fit the different articulations, we convert both the CT data and the atlas into a surface representation. Subsequently, we apply the iterative closest point (ICP) algorithm [8], which minimizes the Euclidean distance between two point sets. The articulated registration is performed by traversing the hierarchical anatomical tree in a top-down manner, optimizing the DoFs specified in each model node. After convergence for a node level, the error criterion is further minimized for the lower node level, yielding a gradually decreasing error function. Depending on the joint type, these DoFs differ per node. However, all registration steps include translations, a varying number of rotations and anisotropic scaling to capture possible differences in size between individuals (see Table 1). The error criterion is minimized with respect to the current node parameters using Levenberg-Marquardt minimization.
x
y
z
729
An exception to this scheme is the spine which is not determined by registration but by binning the bone point set along the longitudinal axis and applying three dimensional region growing, starting from the head-spine connection. The amount of points in each bin increases significantly when the spine-pelvis connection is reached. This initializes the pelvis registration. 2.4. Coarse whole-body alignment To initialize the articulated registration, the mouse model needs to be coarsely aligned to the skeleton segmented from CT i.e. global DoFs have to be removed. Figure 2 shows that the first principal axis of the bone voxels defines the longitudinal body axis (further referred to as z-axis). By binning the bone voxels along the z-axis and computing the Center of Gravity (CoG) for each bin, we derive a 3D curve that enables coarse alignment in the following manner. First, a possible rotation around the longitudinal axis is resolved. Due to the relatively larger asymmetry of the skeleton in the vertical plane (yz-plane in Figure 2) compared to the horizontal plane (assuming that the limbs are placed somewhat symmetrical to the vertical plane), projecting the CoG curve on the xy-plane yields a 2D point set whose first eigenvalue is significantly larger than its second eigenvalue. Thus, the first eigenvector indicates the vertical direction. Second, prone or supine subject position can be derived from the CoG curve between the rib cage and the pelvis. Third, the front and hind part of the subject can be distinguished by assessing the amount of bone along the z- axis where the front shows a peak in bone density due to the skull. Finally, a possible translation along the z-axis is
resolved using a significant maximum of the CoG curve in y around the head-spine connection. An initial scale factor is derived from the relative bone volume of the CT skeleton compared to the atlas bone volume. The initial alignment parameters are stored in the top node of the hierarchical part tree.
3. EXPERIMENTAL RESULTS
To test the registration framework, 3D data volumes were acquired of 12 mice with different postures with a Skyscan 1178 Micro-CT scanner. The original data resolution was 80x80x80 m3. The data was sub-sampled with a factor 4, smoothed and the skeleton was segmented through isodata thresholding. While the coarse structure of all bones that are included in the registration process is retained, all the ribs have been removed by the preprocessing steps to ensure robust matching of the sternum. Since the shoulder blades are very thin, they are not properly represented after the preprocessing steps and are therefore not considered. The resulting skeleton representation is subsequently used for the coarse alignment step as described in the previous section. Two representative registration results on mice with very different postures are shown in Figure 3. Isosurfaces of the segmented skeletons are shown in grey and the phantom skeleton surface in red. The upper and the middle row show the skeletons before and after coarse registration respectively. The last row shows the results after the articulated registration step. The spine is represented by the CoGs of the bins that were used for the 3D region growing.
Figure 3: Two examples (left and right column) of the registration between the model (red) and segmented micro-CT data (grey) before registration (top row), after the coarse alignment step (middle row) and after the articulated registration (bottom row)
730
00 Coarsely aligned skeleton 01 Skull 02 Right part of the pelvis 03 Right hind upper limb 04 Right hind lower limb 05 Right hind paw 06 Left part of the pelvis 07 Left hind upper limb 08 Left hind lower limb 09 Left hind paw 10 Breast bone 11 Right front upper limb 12 Right front lower limb 13 Right front paw 14 Left front upper limb 15 Left front lower limb 16 Left front paw
Figure 4: Decrease of the error criterion (minimum distance from a phantom surface node to a CT skeleton surface node) calculated including all phantom surface nodes (approx. 3000 nodes and approx. 15000 nodes for the CT skeleton surface) for all 12 datasets as the hierarchical tree is traversed (left) and mean and standard deviation of the average error for specific bones from all 12 datasets before and after registration (right)
All 12 cases yielded a similar correct registration between the CT-segmented skeleton and the atlas. The decrease in the registration error for all 12 datasets is plotted in Figure 4 (left). The registration error including all phantom surface nodes decreases from an average of 2.93 ±0.63mm to 0.44 ±0.04mm. Figure 4 (right) shows the averaged bone specific registration result for all 12 datasets. One can clearly see that the lower a bone is located in the hierarchy, the larger the error is before the registration. After registration, there is no relation anymore between error and hierarchical level. The entire registration procedure was implemented in Matlab 7.1 and took less than 2 seconds for the coarse registration and 2-3 minutes for the articulated registration on a standard desktop PC.
4. DISCUSSION AND CONCLUSION We presented a fully automated method for atlas-based segmentation of an entire mouse skeleton, and showed experimental results on real data. The chosen ICP algorithm, applied to the surface representations of CT data and the atlas, proved to lead to a registration result with an average surface node distance of 0.44 ±0.04mm, including all datasets. While in general the registration result is very good, there are some areas where parts of the atlas elements deviate from the real data. This may be caused by non-linear shape differences between the atlas and the data, which can be removed by applying a subsequent non-rigid registration. Also, implementation of a second iteration step that takes the first one as an initialization and/or using data at higher resolution may improve registration accuracy and allows including thinner bones like the shoulder blades or the ribs as well. Though computation time is generally acceptable, it could be improved by defining a volume of interest in the CT data prior to the articulated registration step, after an atlas element is initialized. This can also be realized by removing the model parts that have been already registered from the computations. Also, the “part-of”
concept allows parallel computing and therefore the calculation speed could be increased as well. Based on this work, we will focus on the integration of the atlas soft tissue elements such as the lungs, heart and intestinal organs that can be distinguished in CT data in the registration procedure. The skeleton based registration, combined with the skin surface then serves to constrain a non-rigid registration for the soft tissues, yielding a whole-body segmentation. Also, we expect that this method generalizes well towards other rodents, provided that an anatomical atlas is available.
5. ACKNOWLEDGEMENTS The authors gratefully acknowledge Henk Rozemuller and Maj Petersen for preparing animals for CT scanning, Dr Paul Segars for providing us with the mouse atlas and Elke van de Casteele from Skyscan for providing the Skyscan Micro-CT scanner used in this research.
6. REFERENCES
[1] N. Kovacevic, G. Hamarneh and M. Henkelman, Proc. MICCAI 2003, LNCS, vol 2879, pp 870-877
[2] X. Papademetris, D.P. Dione, L.W. Dobrucki, L.H. Staib and A.J. Sinusas, Proc. MICCAI 2005, LNCS, vol 3750, pp 919-926
[3] A. du Bois d’Aische, M. De Craene, B. Macq and S.K. Warfield, Proc. IEEE Conference on Image Processing 2005, vol 1, pp 21-24
[4] W.P. Segars, B.M.W. Tsui, E.C. Frey, G.A. Johnson, and S.S. Berr, “Development of a 4D digital mouse phantom for molecular imaging research”, Molecular Imaging and Biology 2004, vol 6(3), pp 149-159
[5] http://www.mc.com/tgs [6] M.J. Cook, “Anatomy of the Laboratory Mouse”,
ISBN 0121869504 [7] http://preventdisease.com/fitness/fundament/articles/ty
pes_of_joints.html (01/12/2006) [8] P.J. Besl and N.D. McKay, “A method for registration
of 3D shapes”, IEEE Transactions on Pattern Analysis and Machine Intelligence 1992, vol 14(2), pp 871-881
731