Automated Segmentation of in Vivo and Ex Vivo Mouse Brain Magnetic Resonance Images

Abstract

Segmentation of magnetic resonance imaging (MRI) data is required for many applications, such as the comparison of different structures or time points, and for annotation purposes. Currently, the gold standard for automated image segmentation is nonlinear atlas-based segmentation. However, these methods are either not sufficient or highly time consuming for mouse brains, owing to the low signal to noise ratio and low contrast between structures compared with other applications. We present a novel generic approach to reduce processing time for segmentation of various structures of mouse brains, in vivo and ex vivo. The segmentation consists of a rough affine registration to a template followed by a clustering approach to refine the rough segmentation near the edges. Compared with manual segmentations, the presented segmentation method has an average kappa index of 0.7 for 7 of 12 structures in in vivo MRI and 11 of 12 structures in ex vivo MRI. Furthermore, we found that these results were equal to the performance of a nonlinear segmentation method, but with the advantage of being 8 times faster. The presented automatic segmentation method is quick and intuitive and can be used for image registration, volume quantification of structures, and annotation.

THE VERSATILITY of magnetic resonance imaging (MRI) techniques makes animal MRI suitable for the identification of new disease biomarkers and evaluation of novel diagnostic or therapeutic agents, similar to clinical MRI.^1–3 For the validation of these newly developed techniques for neuroanatomic studies, the MRI volumes are often compared with histologic sections. Therefore, an automatic method to register a single histologic section to a three-dimensional MRI volume registration is desired. However, this is not a trivial task. Several authors tackled the histology to MRI registration problem by using a block head image as an intermediate registration step or manually depicted landmarks to guide the registration.^4–7 In previous work, we found that manually segmenting the MRI volume as a preprocessing step was sufficient to register single two-dimensional hematoxylin and eosin–stained histologic sections to an ex vivo MRI volume of the mouse brain.⁸ The aim of this research was to develop a fully automated segmentation method for in vivo and ex vivo mouse brains, which is fast and accurate enough to be used as an intermediate step for registration, as well as generic enough to be used on different small-animal MRIs.

Automated segmentation of mouse brain MRIs is still very challenging, in contrast to the automated segmentation of human brain MRIs.⁹ Unfortunately, most algorithms developed for the human brain segmentation are not directly applicable to mouse brain images, as Tohka and colleagues recently presented.¹⁰ Although the requirements for the small-animal imaging with MRI are similar, performing volumetric measurement or shape analysis is useful for the annotation purposes or quantitative phenotyping of (transgenic) mouse models. These segmentation problems in mice brain MRI are mostly due to artifacts caused by the MRI scanner, deformations caused by the excision of the brain, and, most importantly, less contrast between brain structures and a lower signal to noise ratio compared with human MRI.

Segmentation of mice brain MRI for clinical practice is nowadays still performed either manually or semiautomated. In the latter case, a presegmented atlas is nonlinearly registered to a new subject and afterward is manually corrected.^11–15 In these studies, no segmentation performance is reported.

A completely automated segmentation method based on nonlinear registration was presented by Rohlfing and colleagues, which reached a segmentation accuracy of 90% overlap with manual contours in bee brain MRI.¹⁶ The method consists of the nonlinear registration to several atlases, which are combined by an expectation-maximization classification method. The success rate of this method is very dependent on the amount of atlases available.

Another promising fully automated segmentation method is based on probabilistic intensity information or intensity patterns of various ex vivo imaging protocols, which are known on forehand.^17–19 With this approach, on average, the automated segmentation had 90% overlap with manually drawn contours. The advantage of this method is that no computation expensive registration methods are required, although the use of various imaging protocols might be time consuming.

In this article, a new, fast, automatic segmentation method is presented that produces segmented images of in vivo and ex vivo mouse brains based on a single atlas, imaged by a single imaging protocol. We first applied a fast affine atlas registration to a template to obtain a rough initial segmentation that was then refined by a clustering algorithm. As already stated by Tohka and colleagues,¹⁰ regular clustering algorithms, such as fuzzy k-means clustering²⁰ and Markov random field models,²¹ fail to segment the volume properly, mainly because there is a lack of contrast between the structures. Therefore, a more specialized clustering algorithm is required, which we present in this study. The presented clustering algorithm combines intensity values, the class labels of the neighboring voxels, and edge information. This information is given to the clustering algorithm by means of a template.

The algorithm is tested on in vivo and ex vivo mouse brain MRI volumes. For both volumes, different structures are segmented based on the visibility and contrast of the structures in the volumes. Furthermore, the performance of the algorithm is validated by comparison with manually drawn expert contours. The in vivo MRI segmentations are also compared with automated segmentations obtained by a nonlinear segmentation method. For this approach, the MRI volume is nonlinearly registered to the atlas by the Demons registration method.²²

Materials and Methods

Experimental Setup

C57Bl6/Jico mice (n = 5) were first imaged by magnetic resonance in vivo on a Bruker 9.4-Tesla scanner using a T₂-weighted multislice spin echo sequence with TR/TE = 6000/35 milliseconds (four averages). The in vivo volume had a matrix size of 256 × 256, with 40 slices, resulting in a resolution of 97.6 × 97.6 × 200 μm per voxel. The total scan time was 102 minutes. Afterward, mice were sacrificed, and the brain was excised from the skull and perfusion fixed with 4% phosphate-buffered paraformaldehyde (PFA). Prior to ex vivo MRI, brains were incubated for 8 hours in 4% PFA containing 12.5 mM gadolinium–tetraazacyclododecanetetraacetic acid (Dotarem, Guerbet, Roissy, France). Ex vivo imaging was performed using a T₁-weighted three-dimensional gradient echo protocol, with TR/TE = 17/7.6 milliseconds and a flip angle of 25°. The total scan time was 10 hours. The ex vivo volume had a matrix size of 256 × 256 × 256 and an isotropic resolution of 78.1 μm per voxel.

Figure 1 displays the pipeline of the presented segmentation algorithm with its two main steps: the registration to an atlas and the clustering. Also, the required input for the algorithm is displayed. In the following, the various brain structures of interest are denoted as classes. The automated segmentation for each image took on a single Pentium-IV 3.4 GHz processor approximately 30 minutes for the ex vivo volume and 15 minutes for the in vivo volume. The difference in calculation time between ex vivo and in vivo volumes is due to the differences in the number of voxels.

Template Creation

In atlas-based segmentation, or model-guided segmentation, a new MRI volume can be segmented if it is registered to an atlas. The atlas contains all previous information on the average volume and spatial organization, which is useful to avoid biologically impossible solutions. The best representative atlas for image segmentation and normalization is an unbiased atlas, which means an atlas that is constructed by averaging scans of multiple subjects in an independent coordinate system and is not dependent on intersubject changes.²³ If insufficient subjects are available for the creation of an unbiased atlas, an approximation can be made as presented by Thompson and Toga, who mapped an unknown brain to a database of normal brains to acquire an accurate segmentation.²⁴

In this study, a limited number of subjects was available, which excluded the possibility of creating an average unbiased atlas. So, we had to work with a template, a single segmented volume that was selected from the in vivo and ex vivo images from the data set. Although an unbiased atlas is desirable, a template would suffice for this purpose. The clustering algorithm is applied after the affine registration, adjusting the initial segmentation until a perfect individual segmentation is reached.

Figure 1.

The segmentation pipeline for the mouse brain segmentation algorithm. The atlas volume is first registered to the new magnetic resonance imaging (MRI) volume, resulting in an initial segmentation. Furthermore, the intensity distributions per class are derived from the atlas, and the edge information from the new MRI volume is extracted. The clustering algorithm is performed in the second step for a final segmentation; therefore, it uses the class statistics, initial segmentation, and edge information as input.

Owing to the differences in intensity values, contrast, and noise between in vivo and ex vivo MRIs, we used both in vivo and ex vivo images of the template. Furthermore, the number of manually segmented structures was also dependent on the visibility of those structures. For the ex vivo atlas, 15 structures were segmented: the cortex, midbrain-hindbrain, cerebellum, olfactory areas, hippo-campal formation, caudate putamen, thalamus, corpus callosum, hypothalamus, fornix system, corticospinal tract, substantia nigra, ventricles, anterior commissure–olfactory limb, and anterior commissure–temporal limb. For the in vivo atlas, 12 structures were segmented: the cortex, midbrain-hindbrain, cerebellum, olfactory areas, hippo-campal formation, caudate putamen, corpus callosum, fornix system, substantia nigra, ventricles, anterior commissure–olfactory limb, and anterior commissure–temporal limb. Coronal, sagittal, and transverse views of the atlas are given in Figure 2.

Registration

Each volume in the data set is affine registered to a manual segmented template, which provides an initial segmentation. Given that the registration is an intermediate step in the segmentation algorithm, a fast and rough registration of the template to the new volume is sufficient. For this purpose, we used a registration algorithm composed of an affine transform with mutual information as image metric, which was optimized by a regular step gradient optimizer^25,26 as implemented in the National Library of Medicine Insight Segmentation and Registration Toolkit (ITK).²⁷

When the registration has finished, the segmentation of the atlas is affine transformed and mapped on the new MRI volume as an initial segmentation. In addition to acquiring an initial segmentation, the atlas was also used to derive prior information on the intensity distribution for each class as input for the clustering algorithm. Furthermore, the initial segmentation is used to remove the skull and surrounding tissue of the in vivo MRI volume, so the clustering algorithm will not be distracted by those.

Figure 2.

The magnetic resonance imaging (MRI) atlas used for in vivo (E, G) and ex vivo (F, H) MRIs with their manual segmentations (A, B, C, and D) and corresponding names and abbreviations. For a better understanding, the abbreviations for the brain structures are in uppercase, whereas the abbreviations for brain tracts are in lowercase.

Edge-Based Clustering

After the atlas-based registration is completed, the clustering algorithm is applied. This clustering is necessary since the affine registration results in an initial segmentation that only accounts for global differences between the new volume and the atlas image. The clustering corrects the segmentation for local changes caused by intersubject variation and deformations in the ex vivo mouse brain caused by the physical excision and preparation of the brain.

For each voxel x in the MRI, a probability is calculated for x to be a member of class c as formulated below. This clustering function uses information on the intensity distribution and information retrieved from the N neighbors of x:

p (c | x, η \in N) = (1 - α) p_{int e n s i t y} (c | x) + α p_{n e i g h b o u r h o o d} (c | x, η \in N)

(1)

with 0 ≤ α ≤ 1

The clustering algorithm evaluates (1) for each class c and assigns x to the class with the highest probability. This process is iterated until convergence is reached. This convergence level is defined as the minimum percentage of voxels changing label in a single iteration. This predefined percentage of voxels has to be set by the user.

The algorithm usually finishes between the 5 and 10 iterations, depending on the convergence threshold set by the user and the quality of the initial segmentation. The latter is provided by the first step of the algorithm, the affine registration, as described in the section above.

The presented algorithm needs three inputs, as can be seen in Figure 1: an initial segmentation, as given by the global atlas-based registration; the intensity distributions per class as derived from the atlas; and the edge information of the new MRI volume. As can be seen in formula (1), the clustering is separated in two main components; the first one, P_intensity(c|x), is based on the intensity distribution of the various classes, and the second part (P_intensity(c|x, n ∈ N)) is based on information retrieved from the neighborhood of voxel x. Knowledge of the intensity distributions is derived from the atlas. The initial segmentation and the edge information are used to calculate the neighbourhood influence.

The weight α is used to tune the algorithm for the various contrast to noise and signal to noise ratios, depending on the imaging protocols of the MRI scanner. If the image volume has very high contrast, the emphasis may lie on the probability from the intensity, so α should be smaller than 0.5. If the data are very noisy, the probability calculated from the intensity is less reliable. In this case, α should be put higher than 0.5 since the edges can still be found correctly by using an anisotropic smoothing filter.

The first part of the clustering algorithm, the P_intensity(c|x), is used to incorporate the intensity distribution of the various classes. It measures the relative distance of each voxel x of the complete volume × to the class mean intensity of each class c (x_c), where the shortest distance has the highest probability of assigning the x to c:

p_{i n t e n s i t y (c | x) = 1 - \frac{(x - \bar{x_{c})^{2}}}{\sum_{x \in X} (x - \bar{x_{c})^{2}}}}

(2)

The second part of the probability function, the P_intensity(c|x, n ∈ N), models the dependency on the neighboring voxels. The influence of the neighbors is weighted by the edge information obtained from the original image since the initial segmentation is usually erroneous near the edges, especially when the segmentation is found by global atlas registration. Therefore, the neighbors that are located inside a structure have more influence than the neighbors located on or close to a (strong) edge. The edges are found by a standard Sobel filter S(x), and afterward, the intensities of the image are scaled to range from 0 to 1. For this algorithm, we use a second-order neighborhood, which means that all voxels located next to x in a horizontal, vertical, and diagonal direction are included in N, thus resulting in a neighborhood of N = 3³ = 27 voxels, including x itself. The n_c in formula (3) symbolizes that for each class c, neighbors can contribute only if they are also classified to c:

p_{n e i g h b o u r h o o d} (c | x, n \in N) = \sum_{n_{c} \in N} \frac{1 - S (n_{c})}{N}

(3)

Validation

For validation purposes, all brains were manually segmented in concordance with two experts, who used the standardized mouse brain atlas of the Laboratory of Neuro Imaging at the University of California²⁸ as guidance. The structures that were selected for manual and automated segmentation were selected by the experts based on their visibility in MRI. The results of the automated segmentation method were validated by comparing them with the manual segmentations of experts by means of the kappa index.²⁹ The kappa index is a measure that represents a ratio of the amount of overlap to the total number of voxels of an automatically segmented brain structure V_a and a manually segmented brain structure V_m:

k = \frac{2 (V_{a} \cap V_{m})}{V_{a} + V_{m}}

(4)

This measure is robust to changes in volume size and therefore very suitable to compare the automated and the manual segmentation. A kappa of 1.0 indicates total overlap of two volumes, whereas a kappa of 0.0 shows no overlap at all. In an interobserver study, it was shown that an automated segmentation method performs as well as human observers if kappa indices between 0.7 and 1.0 can be achieved.¹⁷

An overall validation score of the algorithm is obtained by averaging the kappa indices per structure for all volumes in the data set.

Results

Automated Segmentation

For the segmentation of the ex vivo images, the algorithm used on average six iterations to reach the threshold when less than 0.5% of the voxels changed label, whereas for the in vivo segmentation, only three iterations were needed for convergence to a 5% threshold. The different settings for the convergence threshold are a consequence of the different voxel sizes of the mouse brain in the ex vivo and in vivo images, which are, respectively, 127,655 voxels and 833,800 voxels. The automated segmentation for each image took on a single Pentium-IV 3.4 GHz processor approximately 20 minutes for the ex vivo volume and 15 minutes for the in vivo volume. The difference in calculation time is due to the differences in the number of voxels.

The average kappa indices are calculated for all automatically segmented structures and displayed in Table 1. Also given are the volumes of the segmented structures in voxels. As stated in the previous section, an automatically segmented structure with a kappa index larger than 0.7 can be assumed to be segmented with reasonable accuracy. If we consider the in vivo automated segmentation, the algorithm reached a satisfying result for 7 of the 12 structures with an average kappa index of 0.7. In the case of the ex vivo segmentations, 12 of 15 structures are correctly segmented with an average kappa index of 0.85.

Table 1.

Sizes (in mm³ and in voxels) for Several Brain Structures

	In Vivo Segmentation			Ex Vivo Segmentation
Structure Name	Volume (mm³)	Number of Voxels	Kappa Index	Volume (mm³)	Number of Voxels	Kappa Index
Cortex	157.07	82,443	0.884 ± 0.005	124.11	260,280	0.884 ± 0.001
Midbrain-hindbrain	94.68	49,696	0.918 ± 0.09	76.29	159,995	0.924 ± 0.027
Cerebellum	51.92	27,250	0.896 ± 0.005	50.93	106,801	0.904 ± 0.014
Olfactory areas	24.95	13,098	0.822 ± 0.010	22.57	47,341	0.739 ± 0.230
Hippocampal formation	21.67	11,377	0.826 ± 0.014	22.06	46,268	0.916 ± 0.019
Caudate putamen	20.15	10,577	0.813 ± 0.020	20.20	42,353	0.906 ± 0.021
Thalamus	—	—	—	26.35	55,260	0.888 ± 0.078
Corpus callosum	13.12	6,887	0.578 ± 0.028	18.72	39,260	0.808 ± 0.006
Hypothalamus	—	—	—	9.96	20,884	0.86 ± 0.072
Fornix system	4.85	2,548	0.53 ± 0.038	6.14	12,886	0.721 ± 0.038
Corticospinal tract	—	—	—	4.38	9,183	0.709 ± 0.033
Substantia nigra	1.93	1,012	0.687 ± 0.021	1.26	2,634	0.729 ± 0.181
Ventricles	8.43	4,427	0.778 ± 0.039	2.66	5,579	0.508 ± 0.051
Anterior commissure-olfactory limb	0.52	275	0.425 ± 0.216	1.45	3,047	0.557 ± 0.039
Anterior commissure-temporal limb	0.36	190	0.322 ± 0.110	0.22	461	0.401 ± 0.062

For the separate structures, the in vivo and ex vivo segmentation results are given in average kappa indices and standard deviation. The missing values of the in vivo segmentation column represent structures for which no proper expert segmentation could be obtained and thus were excluded from the atlas and automated segmentation.

These results imply that the automated segmentation method for ex vivo MRIs is comparable to manual segmentations, as Ali and colleagues showed with an intraobserver study,¹⁷ which reach an average kappa of 0.85. The automated segmentation method from Sharief and colleagues outperformed the presented method with an overall kappa index of 0.95.19 However, their method is applicable only to ex vivo MRI since they use various imaging protocols, whereas the presented method is also applicable to in vivo MRI.

If the number of voxels is compared to the final kappa index, it can be seen that the performance of the algorithm decreases with the size of the structure to be segmented. This holds especially for the brain tracts included in the segmentation algorithm: the corpus callosum, corticospinal tract, and anterior commissures.

To illustrate the differences in segmentation for the in vivo and ex vivo images, an automatically segmented slice and a manually segmented slice are displayed in Figure 3. Most structures have an overall better segmentation result in the ex vivo images, owing to a better contrast to noise ratio and higher resolution. The effect of these parameters is clearly visible if one considers, for example, the corpus callosum. However, some brain structures suffer from major deformations caused by the excision of the brain, impairing an accurate automatic segmentation of ex vivo. Examples are the olfactory areas, which are easily damaged during excision, or the ventricles, which often collapse postmortem and therefore have a smaller volume, leading to worse segmentation results for the ex vivo images. In Figure 3, one can see clearly differences in proportion for the ventricles in in vivo and ex vivo images.

In Figure 4, we presented the kappa indices of the ex vivo segmentations after the first step of the algorithm (the affine atlas-based registration) and its second step (the clustering algorithm). This figure shows that only for the three brain tracts, the fornix system and anterior commissures, is a decrease in kappa index obtained after the clustering is performed. For the anterior commissures, this decrease is even found to be significant. The kappa index increases for all other structures, and although this increase seems unimportant and small in the figure, we found a significant increase in kappa index for all structures except the cortex, midbrain-hindbrain, and caudate putamen. For these three structures, the initial segmentation is already quite accurate, leading to minor adjustments by the clustering algorithm. These minor corrections may not be a significant improvement but are still important since these small changes are actually corrections for the intersubject variations. The structures, for which a significant improvement in segmentation was found, are the structures that have a less accurate initial segmentation compared with the cortex, midbrain-hindbrain, and caudate putamen and therefore need more correction by the clustering algorithm.

Figure 3.

A visual comparison between the manual (M) and automated (A) segmentation for in vivo magnetic resonance imaging (MRI) (top) and ex vivo MRI volumes (bottom). The color coding of the various classes correspond to the legend as given in Figure 2.

Nonlinear Atlas-Based Segmentation Results

As commonly acknowledged, the nonlinear registration to an atlas is the most accurate way to segment brain volumes. To compare the performance of our algorithm with atlas-based nonlinear segmentation, we used the symmetric Demons algorithm as implemented in ITK. This nonlinear segmentation is based on a thermodynamic concept of diffusion. Having two images to match, the main idea is to consider object boundaries in one image as semipermeable membranes and let the objects of the other, deformable or moving image diffuse through these membranes and deform the image such that it fits the mold of the fixed image. In medical images, isointensity contours are closely related to object shapes since intensity reflects the tissue properties.²²

The same procedure was followed as for the presented segmentation method; the same in vivo image was used as an atlas, as shown in Figure 2, whereas the segmentation algorithm was evaluated on the other in vivo images. Before applying the Demons algorithm, the in vivo images were affine registered by the same algorithm as used for the atlas-based registration of the newly presented method.

The average kappa index is retrieved by comparing the results from the automated segmentation with the results of the manual segmentation of the in vivo mouse brain volumes.

The results of the nonlinear registration to the atlas can be found in Figure 5. As can be seen in the figure, the results of the algorithm are comparable with the results of the Demons algorithm, where the clustering method reached convergence in 10 minutes and the Demons algorithm reached convergence in 2 hours on the same computer.

Figure 4.

The increase in kappa index between the two steps of the presented algorithm for the ex vivo segmentation results. The light gray bar denotes the average kappa index after the affine atlas-based registration step, whereas the dark gray bar displays the average kappa index after the clustering step. Furthermore, the standard deviations are given for each bar to indicate the robustness of the algorithm. The abbreviations of the various structures are explained in Figure 2.

Figure 5.

The average kappa indices of the Demons algorithm (light gray bars) compared with the average kappa indices of the presented method (dark gray bars) for the in vivo mouse brain segmentation. Furthermore, the standard deviations are given for each bar to indicate the robustness of the algorithm. The abbreviations of the various structures are explained in Figure 2.

Discussion and Conclusion

As described in the Results, the algorithm was found to segment the larger brain structures, for example, the cortex or cerebellum, correctly. For these structures, the algorithm performs equally to the kappa indices from literature.¹⁷ The results are also comparable for one of the most challenging structures in the ex vivo brain, the corpus callosum. This structure is challenging to segment because it is assumed to be large, with an average of 37,990 voxels (18.11 mm³), but its flat and thin shape resembles more a small structure.

The algorithm has a high segmentation performance on structures that are affected by noise or deformations and have low contrast. This performance can be obtained by the use of neighborhood voxels in combination with the edge information of the unknown brain image. A drawback of this method is, however, that the algorithm has less accurate segmentation on structures with a width of only one voxel. This can be seen in Figure 4, where the three brain tracts, the fornix system and anterior commissures, show a decrease in kappa index after the clustering is performed. On the one hand, this segmentation error is due to the use of a single image as atlas for the segmentation: small intersubject variations result in a misregistration for the smaller structures in the brain, in such a way that there is no overlap and thus no seed point for the algorithm to segment this structure. On the other hand, if structures of one voxel width are present at the correct location, it might be that the neighboring structures have more intensity similarity and a higher probability based on more labeled voxels, so that the small structure will be assigned to an incorrect label.

By comparing the in vivo with ex vivo segmentations, the algorithm returns a better segmentation for the ex vivo images, except for the olfactory areas and ventricles. The superior segmentation of the ex vivo segmentation can be explained by the lower resolution of the in vivo volume, resulting in more structures with a single voxel width. As mentioned above, the algorithm encounters more difficulties when segmenting structures with a limited number of voxels. The anisotropic voxel size of the in vivo volume has no effect on the quality of the registration. This is shown by Table 1, which gives the number of voxels per structure. The superior segmentation of the olfactory areas and ventricles in the in vivo volumes can be explained by the local deformations occurring during the extraction of the brain from the skull. The olfactory areas are very loosely attached to the brain and in most cases were damaged or completely removed in the procedure, whereas the ventricles collapse when the brain is fixed. The latter is shown in Table 1, where the volumes in cubic millimeters are given for all structures. Although the segmentation algorithm can compensate for these changes to some level, these deformations still cause some errors.

The boundaries of some brain structures, for example, the transition of the midbrain-hindbrain to the thalamus, are difficult to determine owing to little contrast differences in the image between the structures. For these structures, the manual segmentation is also subjective and differs for each mouse brain. In this study, two experts were used to validate the manual segmentation and obtain a more objective segmentation. However, more information on the user variability is needed for these structures before some conclusions can be drawn on the quality of the automated segmentation of these structures.

The automated segmentation of the structures with poor edge information is very dependent on the atlas since the clustering is guided by the prior information given by the manual segmentation of the atlas. For these structures, the main differences between the automated segmentation and the manual segmentation are the differences in the transformed expert segmentation of the atlas mapped on the image on the one hand and the manual segmentation of the new image as validation on the other hand. This raises the need for an average atlas that also includes intervariability and intravariability information for these structures, as already performed on human brains.³⁰

We found that the presented segmentation method performs similarly to the nonlinear segmentation method. If compared with the expert segmentations, the performance of the presented method is more consistent. Manual segmentation is still considered the most reliable method, although the intraobserver variation is on average higher than in automated algorithms. This is due to different interpretations of the various structures, as well as the tiredness of the observer. Therefore, an automated segmentation algorithm not only reduces the amount of time needed to segment but also improves the objectivity of the segmentation. In particular, when there is good contrast between structures, the automated segmentation algorithm will return a good and objective segmentation, which is also repeatable.

The algorithm has one limitation caused by the extraction of previous knowledge on intensities and edges from the atlas. If the imaging protocols differ, incorrect intensity distributions per brain structure are derived from the atlas and do not represent the intensity distribution per structure. Given that these distributions are used to guide the clustering, the clustering will result in an incorrect segmentation. So, it is required that the atlas is either acquired with the same imaging protocol as the image data set or has to be preprocessed by some intensity transform to map the intensities on the protocol of the new image data set. In practice, the last method is the most likely choice, although errors made in the intensity mapping will induce errors in the segmentation of the volumes. If an atlas—or example segmentation—can be obtained, it is more likely that a better segmentation result is reached.

Future work will also include a study on the segmentation of other types of MRI. We will investigate the performance of this segmentation method for other images since no specific brain tissue information is used; consequently, all of the posterior information for the clustering is derived directly from the atlas.

In summary, the presented method is a quick and promising segmentation method for mouse brain images, especially when major deformations of the tissue are absent. The smaller, local deformations in the brain tissue are corrected by the adapted clustering algorithm as a complement of the linear registration. This collaboration of both segmentation algorithms results in a quick and accurate segmentation method for in vivo and ex vivo mouse brain MRI, despite its low signal to noise ratio and artifacts. Finally, since no previous information has been used in this segmentation algorithm, this algorithm is highly generic and can be applied to various images without any difficulties.

Conclusion

The main objective of this studyis to find a quick and fully automated segmentation method of in vivo and ex vivo mouse brain MRIs, which can be used in the registration of mouse brains to other modalities. The presented method consists of an affine atlas-based registration combined with an edge refining clustering algorithm, where the clustering is supplemented by edge information and statistical information derived from the anatomic atlas. It is shown that the addition of the clustering algorithm improves the segmentation and is able to compensate for some nonlinear deformations in the ex vivo mouse brain. Whereas fully automated and highly accurate segmentation methods for in vivo and ex vivo mouse brains are extremely time consuming, for example, by nonlinear registration, the presented method is quick but accurate enough for the segmentation of the principal structures needed for the registration.

Footnotes

Acknowledgments

Financial disclosure of authors: Funded by the Dutch BISK. We would like to thank the reviewers for their comments and advice on the manuscript, the Center for Medical Systems Biology, Leiden University, the Netherlands, for its cooperation, and the Dutch BISK for the funding.

Financial disclosure of reviewers: None reported.

References

Anderson

Frank

. MRI of mouse models of neurological disorders. NMR Biomed 2007;20:200–15.

Turnbull

Mori

. MRI in mouse developmental biology. NMR Biomed 2007;20:265–74.

Chen

Kovacevic

Lobaugh

. Neuroanatomical differences between mouse strains as shown by high-resolution 3D MRI. Neuroimage 2005;29:99–105.

Meyer

Moffat

Kuszpit

. A methodology for registration of histological slide and in vivo MRI volume based on optimizing mutual information. Mol Imaging 2006;5:16–23.

Breen

Lazebnik

Wilson

. Three-dimensional registration of magnetic resonance image data to histological sections with model-based evaluation. Ann Biomed Eng 2005;33:1100–12.

Geurts

Pouwels

. Cortical lesions in multiple sclerosis: combined postmortem MR imaging and histopathology. AJNR Am J Neuroradiol 2005;26:572–7.

Bardinet

Ourselin

Dormont

. Co-registration of histological, optical and MR data of the human brain. In: Dohi

Kikinis

, editors. LNCS medical image computing and computerassisted intervention-part I. Tokyo, Japan: Springer Berlin/Heidelberg. 2002. p. 548–55.

Scheenstra

AEH

Dijkstra

van de Ven

RCG

. Automated registration of histology sections with ex vivo MRM volumes. In: Proceedings of the International Society for Magnetic Resonance in Medicine (ISMRM). Seattle: 2006. p. 2012.

Gholipour

Kehtarnavaz

Briggs

. Brain functional locatization: a survey of image registration techniques. IEEE Trans Med Imaging 2007;26:427–51.

10.

Tohka

Krestyannikov

Dinov

. Genetic algorithms for finite mixture model based voxel classification in neuroimaging. IEEE Trans Med Imaging 2007;26:696–711.

11.

Smith

Hof

. In vivo 3D digital atlas database of the adult C57BL/6J mouse brain by magnetic resonance microscopy. Front Neuroanat 2008;2:1–10.

12.

Hof

Grant

. A three-dimensional digital atlas database of the adult C57BL/6J mouse brain by magnetic resonance microscopy. Neuroscience 2005;135:1203–15.

13.

Kovacevic

Henderson

Chan

. A Three-dimensional MRI atlas of the mouse brain with estimates of the average and variability. Cereb Cortex 2005;15:639–45.

14.

Bock

Kovacevic

Lipina

. In vivo magnetic resonance imaging and semiautomated image analysis extend the brain phenotype for cdf/cdf mice. J Neurosci 2006;26:4455–9.

15.

Lau

Lerch

Sled

. Longitudinal neuroanatomical changes determined by deformation-based morphometry in a mouse model of Alzheimer's disease. Neuroimage 2008;42:19–27.

16.

Rohlfing

Russakoff

Maurer

. Performance-based classifier combination in atlas-based image segmentation using expectation-maximization parameter estimation. IEEE Trans Med Imaging 2004;23:983–94.

17.

Ali

Dale

Badea

Johnson

. Automated segmentation of neuroanatomical structures in multispectral MR microscopy of the mouse brain. Neuroimage 2005;27:425–35.

18.

Allan Johnson

Ali-Sharief

Badea

. High-throughput morphologic phenotyping of the mouse brain with magnetic resonance histology. Neuroimage 2007;37:82–9.

19.

Sharief

Badea

Dale

Allan Johnson

. Automated segmentation of the actively stained mouse brain using multi-spectral MR microscopy. Neuroimage 2008;39:136–45.

20.

Duda

Hart

Stork

. In Pattern classification. 2nd ed.

John Wiley & Sons, Inc.

New York; 2001. p. 528–30.

21.

Geman

. Stochastic relaxation, Gibbs distributions, and the Bayesian restoration of images. IEEE Trans Pattern Anal Mach Intell 1984;6:721–41.

22.

Thirion

. Image matching as a diffusion process: an analogy with Maxwell's demons. Med Image Anal 1998;2:243–60.

23.

Joshi

Davis

Jomier

Gerig

. Unbiased diffeomorphic atlas construction for computational anatomy. Neuroimage 2004;23:S151–60.

24.

Thompson

Toga

. Detection, visualization and animation of abnormal anatomic structure with deformable probabilistic brain atlas based on random vector field transformations. Med Image Anal 1997;1:271–94.

25.

Maintz

JBA

Viergever

. A survey of medical image registration. Med Image Anal 1998;2:1–36.

26.

Hill

DLG

Batchelor

Holden

Hawkes

. Medical image registration. Phys Med Biol 2001;46:R1–45.

27.

Yoo

Ackerman

Lorensen

. Engineering and algorithm design for an image processing API: a technical report on ITK - The Insight Toolkit. In: Westwood

, editors. Proceedings of Medicine Meets Virtual Reality. Amsterdam: ISO press; 2002. p. 586–92.

28.

MacKenzie-Graham

Lee

Dinov

. A multimodal, multidimensional atlas of the C57BL/6J mouse brain. J Anat 2004;205:71–149.

29.

Kier

. A shape index from chemical graphs. Quantitative Structure Activity Relationship 1985;4:109–16.

30.

Lorenzen

Prastawa

Davis

. Multi-model image set registration and atlas formation. Med Image Anal 2006;10:440–51.