To report the detection of structural and functional biological changes in living animals using small animal in vivo MRI that complements traditional ex vivo histological techniques. We report the development and validation of the application of large deformation high dimensional mapping (HDM-LD) segmentation for the hippocampus in the rat.
Materials and Methods
High resolution volumetric T2 weighted MRI images were acquired at 4.7 Tesla from six male in-breed nonepileptic Wistar rats. Two HDM-LD segmentations of the hippocampus (automated 1 and automated 2) were compared with the manual segmentations of two investigators who independently segmented the hippocampi (manual 1 and manual 2).
The mean overlap for the hippocampi between automated 1 and automated 2 for the right hippocampi was 94.4% (SD 1.0) and for the left hippocampi was 94.3% (SD 2.5), while the mean overlap between automated 1 and manual 1 for the right hippocampi was 91.4% (SD 1.3) and for the left hippocampi was 91.9% (SD 1.4). Mean values for absolute differences for comparisons of all the segmentations were the following: automated 1 versus automated 2, 3.2% (SD 1.0); manual 1 versus manual 2 6.82% (SD 5.22); automated 1 versus manual 1 13.0% (SD 1.8).
THE HIPPOCAMPUS PLAYS a central role in many different neuropsychiatric diseases, including epilepsy (1), schizophrenia (2), and Alzheimer's disease (3). This has led to extensive study of the hippocampus in animal models of neuropsychiatric disease (4, 5). The recent rapid advances in the technology and availability of dedicated small animal in vivo imaging systems, specifically MRI, allows the detection of structural and functional biological changes in living animals, and complements traditional ex vivo histological techniques. A particular advantage of this approach is the ability to perform serial imaging of animals, increasing the power to track the ontogeny of the changes, their relationship to behavioral aspects of the disease, and potentially, the effect of interventions. The findings on postmortem histological examination can be correlated with the imaging results, and thus the respective advantages of the temporal versus spatial resolution of the two methods can be harnessed.
In both humans and animals, past investigators have used manual segmentation of the hippocampus on MRimages to determine hippocampal volumes (4, 6, 7). While the sensitivity for detecting hippocampal asymmetry of manual segmentation as compared to visual inspection is greater in some investigatory settings (8), manual segmentation is time consuming and requires expertise in the details of hippocampal anatomy for accurate segmentations. The difficulty in manual segmentations lies in the subjective interpretations of anatomical variations. The emerging field of Computational Anatomy (9) founded on general pattern theory provides tools and a frame work for accommodating and studying this variability (10–12). Large deformation high dimensional mapping (HDM-LD) is based on principles of computational anatomy, and uses the power of computer assisted shape recognition to identify general patterns within image data. An electronic reference atlas, which can be deformed according to the properties of a viscous fluid, is used to study a hippocampus of interest according to several defined landmarks. Computational anatomic techniques produce three-dimensional (3D) surface representations of the hippocampus with resolution at a subvoxel level, enabling visualization of details of hippocampal surface anatomy (13). HDM-LD extends analysis of longitudinal MRI imaging beyond the volumetric to subregional and morphometric analyses (1).
Several past studies have demonstrated the precision of deformation-based hippocampal segmentation of the hippocampus in humans (14, 15). Haller et al (16) described a deformation-based hippocampal segmentation technique and verified the precision of this technique in normal subjects and schizophrenic patients. Hogan et al showed reproducibility of 92% overlap of sequential deformation-based hippocampal segmentations in subjects with epilepsy and mesial temporal sclerosis (15). HDM-LD identifies subtle neuroanatomical and longitudinal changes in diseases such as epilepsy, Alzhiemer's disease, depression, and schizophrenia, that are not detectable with standard volumetric methods (3, 17–19).
Using a high resolution volumetric acquisition MRI protocol, we have developed a technique for applying HDM-LD to perform segmentations of the rat hippocampus. To verify the precision of HDM-LD segmentation in the rat hippocampus, we compared HDM-LD segmentations with the manual segmentations of two investigators who independently segmented the hippocampi. These studies for validation of the HDM-LD technique were performed in six in-bred nonepileptic Wistar rats. We hypothesize that MRI-based HDM-LD segmentations of the rat hippocampus will be at least as precise and reproducible as manual hippocampal segmentations.
MATERIALS AND METHODS
In-bred 7-week-old male nonepileptic Wistar rats from the breeding colony at the involved hospital were used. All procedures on live animals were approved by the Ethics Committees of the involved institutions.
MR Image Acquisitions
MR images were acquired on a 4.7 Tesla (T) Bruker Biospec 47/30 Avance small-animal spectrometer (Ettlingen, Germany) using a shielded-gradient set (Bruker Biospec) appropriate for rats. Radiofrequency (RF) pulse transmission and MR data acquisition were performed using a 72-mm inner-diameter birdcage coil (Bruker Biospec) optimally tuned to the 1H frequency. T2-weighted axial structural images were obtained contiguously through the entire brain, using a fast spin-echo sequence (acquisition time, 298 s; repetition time, 3.1 s; echo time [TE], 67.5 ms; matrix, 256 × 256; numerical aperture, 3; field of view, 6 cm × 6 cm; voxel dimensions 0.234375 mm × 0.234375 mm × 0.5 mm; rapid acquisition with relaxation enhancement (rare) factor, 8). Animals were scanned under anesthesia in the prone position in a custom-built Plexiglas holder to ensure consistent positioning of the animal. Rats were anesthetized with 5%isoflurane in 1:1 air/oxygen and then maintained on 1.5–2.5% isoflurane for the remainder of the experiment. Images were collected using Paravision 3.0 (Bruker Biospec).
Postacquisition MR Image Processing
MRI acquisitions from six rats were used. One MRI data set, selected randomly, was used to construct an “atlas” image. The other five MRI data sets were used as “target” images for deformation-based hippocampal segmentation, as well as for manual hippocampal segmentation. Analyze AVW version 7.0 (Rochester, MN) was used as the software package for initial image processing. MR images were cropped and converted to an isotropic voxel dimension of 0.1171875. The intensities of the MRI dataset used for the “atlas” image were adjusted, using visual inspection, for optimal contrast of gray and white matter structures, and converted from 16-bit to 8-bit intensity ranges. The range of inclusive intensities was 3,750–16,000 for the 16-bit images, which were converted to a range of 0–255 for the 8-bit images. The intensity ranges of the additional 5 target images were then also converted from 16-bit to 8-bit images, using the same inclusive intensity ranges as the “atlas” image.
For construction of the atlas hippocampal surface image, the hippocampus was segmented, with verification in the coronal, sagittal, and horizontal planes, following the neuroanatomical boundaries of the hippocampus as described by Paxinos and Watson (20) by a single investigator. The final atlas segmentation was reviewed by a second expert segmenter for verification of the surface anatomical boundaries of the hippocampi. The segmentations were saved as binary files.
Subsequent image processing steps were performed using proprietary Linux-based software. To create the hippocampal atlas, hippocampal surface images were created by rescaling the intensity maps of the binary files, with a subsequent single iteration of a surface “smoothing” function. For deformation-based segmentations, the five “target” MR data sets underwent preprocessing in preparation for the mapping algorithm. This consisted of two steps: global landmarking and hippocampus-specific landmarking. Figure 1 shows a diagram outlining the steps of landmarking and deformation mapping of the hippocampal surface of the target MRI.
Landmarking provided an initial condition for the intensity-matching algorithm by roughly aligning the rat and atlas scans. The first step in landmarking was identifying global landmarks which scale and align the atlas brain to the target brain, relying principally on alignment of the borders of the hemispheres, using global landmarks. The second step was individually landmarking each hippocampus. This was done by first identifying and landmarking the septal and temporal pole of each hippocampus (21), which specifies an axis for frame of reference for landmarking of each hippocampus. Then, four landmarks were identified on five cross-sections equally spaced along this axis. The landmarks were placed on the medial, lateral, superior, and inferior border of the hippocampus on each cross-section. Figure 2 shows an illustration of the 4 landmarks in one of the oblique planes. The oblique plane was generated by the algorithm after landmarking of the temporal and septal pole of the hippocampus.
Images and landmarking data were then integrated into another Linux-based software program. Within this program, the mapping algorithm used a coarse-to-fine procedure for generating a transformation field from an atlas reference MR to the target MR. The “coarse” aspect of the procedure relied on the landmark information provided by the expert segmenter to provide an initial low-dimensional coregistration of atlas and target images (22). The landmark information was provided in the form of the global and hippocampus-specific landmarks described in the previous section which was used to derive a coarse manifold transformation (23) from the atlas to the target images. Therefore, the initial “course” step principally consisted of aligning landmarks between the atlas and target images.
Having completed the course first step in the transformation, the volumes were roughly aligned and attention focused on the fine features of the substructures. The “fine” procedure involved the next two steps. The second step was to solve the registration problem using a linear elastic basis formulation and the full volume data, as previously described (11, 24). This was fully automatic and only driven by the volume data itself. The 3D whole brain maps corresponded to the maximizer, whose variation solution corresponded to a solution of a nonlinear partial differential equation (PDE), consisting of between 107 and 108 parameters. The third and final step of the algorithm was to solve the nonlinear PDE corresponding to the Bayesian maximizer associated with the fluid formulation at each voxel of the full volume (10, 25, 26).
The surfaces that were first created using the marching cubes algorithm tended to have highly nonuniform triangles. Refinement toward a more uniform surface was needed so that other computations on the surfaces could be performed, and this procedure was given by Joshi et al (27). Briefly, an elastic energy function was defined on the original surface vertices based on the distance between neighboring vertices. Minimizing this energy function would result in a surface whose vertices are more uniformly distributed across the entire surface, therefore giving the smoother appearance. Because the minimization involving spatial derivatives was performed on the local tangent planes at each vertex, the overall shape characteristic was preserved.
For the deformation-based segmentations, each target data set was completely preprocessed with landmarking twice by a single investigator, at an interval of one month, and processed using the deformation-based algorithm to generate two separate sets of deformation-based segmentations, designated as A1 and A2.
Two investigators independently performed manual segmentations of the hippocampi of the five target MR datasets. The segmentations were designated as M1 and M2.
Comparisons between two segmentations were made by computing the percentage of overlap of voxels, as in previous studies (15). One segmentation was designated the reference (R) and the other the study (S) segmentation that we compared against the reference. The percentage of overlap was computed as the number of overlapping segmented voxels between the two segmentations divided by the total number of segmented voxels in the study, that is, (R intersect S)/S × 100. We used the manual segmentations of one investigator (Y.R.L.) as the reference segmentations, and the first series of deformation segmentations (A1) as the comparison segmentations.
Table 1 shows comparisons of three segmentations: two automatic segmentations (automatic segmentation 1 [A1] and automatic segmentation 2 [A2]) and one manual segmentation (manual segmentation 1 [M1]). The mean percentage overlap for the hippocampi between A1 and A2 for the right hippocampi was 94.4 (SD 1.0) and for the left hippocampi was 94.3 (SD 2.5), while the mean percentage overlap between A1 and M1 for the right hippocampi was 91.4 (SD 1.3) and for the left hippocampi was 91.9 (SD 1.4).
Table 1. Comparison of Percentage Overlap of Voxels Between Segmentations
First Versus Second Automatic Segmentations (A1 vs. A2)
First Automatic Versus First Manual Segmentations (A1 vs. M1)
Table 2 shows volume measurements based on the four segmentations. Mean values for absolute percentage differences for comparisons of all the segmentations were the following: A1 versus A2, 3.2 (SD 1.0); M1 versus M2, 6.82 (SD 5.22); A1 versus M1, 13.0 (SD 1.8).
Table 2. Volume Measurements and Between-Method Hippocampal Volume Differences
Percentage Difference Between Methods (A1 vs. M1)
A1 Volume (mm3)
A2 Volume (mm3)
Absolute Percentage Difference
M1 Volume (mm3)
M2 Volume (mm3)
Absolute Percentage Difference
Figure 3 shows coronal sections of a randomly selected rat MRI (MRI 2) taken at approximately 1-mm intervals. Figure 3A–E progresses from a rostral to caudal direction, with the white shading representing the manual segmentation, and the line representing the automated segmentation. The major differences in the segmentation are represented in Figure 3C, which represents the region where the axis of the hippocampus changes in direction. Figure 3F–J is a replication of the MRI sections shown in Figure 3A–E, without superimposed automated segmentations.
Figures 4–6 show illustrations of the HDM-LD surface images of hippocampi from a single rat MRI. Figure 4 shows a 3D representation of the surface of the hippocampi (Fig. 4D), with associated 2D MR images in the coronal (Fig. 4A), horizontal (Fig. 4B), and sagittal (Fig. 4C) planes. In Figure 4D, the hippocampi are viewed from a rostral perspective. The red marker is placed on the septal pole of the right hippocampus.
Figure 5 shows a 3D view of the hippocampi from a caudal perspective, viewing the surface from the opposite direction as presented in Figure 4.
Figure 6 shows a 3D view of the hippocampi, with the surfaces rotated to show the lateral surface of the left hippocampus. This view serves to show the “C” shape of the left hippocampus.
Past techniques used to quantitate the volume and shape of the hippocampus from MR images of rats have focused on defining structures in a single 2D plane for volumetric measurements (4, 7). Using HDM-LD with general pattern matching, anatomic structures can be segmented using global shape models. By representing the typical structures by means of the construction of templates, and their variability by the definition of transformations applied to the templates, MR images of the hippocampus may be semiautomatically segmented (28). HDM-LD offers advantages over 2D plane reconstructions by providing 3D representations of the involved segmented structure, and allowing mathematical comparisons of surface structure between different images (29).
There has been extensive validation of MRI-based HDM-LD techniques in human subjects (14–16). However, HDM-LD techniques have not previously been applied in the rat brain. The relatively smaller size of the rat brain, and resultant limitations in image resolution of rat brain MRI studies are issues in applying MRI-based HDM-LD. However, with improvements in MRI rat brain imaging, and our current protocol of acquiring high-resolution volumetric images on a 4.7T animal MRI scanner, application of HDM-LD in the rat has become a practical proposition.
The rat hippocampus is a good target for HDM-LD, as it is relatively well demarcated on our MRI acquisitions, and is one of the relatively larger structures in the rat brain. The 3D surface anatomy, however, is relatively complex, with the long axis of the structure turning at an approximately 90 degree angle between the septal and temporal poles of the hippocampus, and therefore running in relatively perpendicular planes at either end of the structure, which is best illustrated in Figure 6. The hippocampus extends from the basal forebrain, over the diencephalon, and caudoventrally into the temporal lobe. The septal pole is located dorsally and rostrally, while the temporal pole is located caudally and ventrally (21). There remain differences in terminology for rat hippocampal structures, while the septal pole is also referred to as the dorsal pole, and the temporal pole is also referred to as the ventral pole (30). The long axis of the hippocampal formation is referred to as the septotemporal axis.
In this study, we show that the results of HDM-LD segmentations and manual segmentations of the rat hippocampus showed good correlations with percentage overlap of structures. The average overlap of the two separate automatic segmentations (A1 and A2) as well as the automatic (A1) and manual (M1) segmentations (Table 1) was better than 90% average overlap for both right and left hippocampi. This helps to verify that the segmentations occupy the same 3D space. Overlap results between segmentations are comparable to those from HDM-LD studies performed in human subjects (15). Volume comparisons (Table 2) show slightly lower average absolute percentage volume differences between the deformation segmentations compared to the manual segmentations. This finding would suggest that the reproducibility of volume measurements using HDM-LD is comparable, if not slightly better, than repeated manual segmentation measurements. However, there are relatively larger volume differences in direct comparison of the deformation and manual segmentation (A1 versus M1), with the manual segmentations showing a consistently larger volume. Figure 1 shows a comparison of the deformation and manual segmentation. Although overlap of the segmentation is good (as indicated in the data of Table 1), there are regions at the margins of the hippocampal borders where there are greater discrepancies in the two measurements. On the MR images, the hippocampal borders as the hippocampus extends ventrocaudally into the temporal lobe (as in Fig. 3C) were often the most difficult to define due to lack of clearly defined boundaries with surrounding structures. As in past validation studies with HDM-LD (15), there is difficulty in defining hippocampal borders, with either the deformation or manual segmentation technique, where hippocampal boundaries are most poorly defined on the MRI images. In general, our results are most consistent with good overlap between the segmentations, showing that both sets of segmentations cover similar regions, but with the manual segmentations being slightly larger, likely related to slightly greater volume inclusion of the manual segmentations at the margins of the segmentations.
The relative size of the rat hippocampus also likely plays a role in our validation results. Past studies have shown greater percentage error when segmentations involve smaller volumes (31). In a study of progressive multiple sclerosis MR lesions, Goodkin et al found a coefficient of variation for three successive lesion measurements inversely related to the lesion area, ranging from 22.6% for lesion less than 0.67 cm2 to 12.1% for larger lesions (31). Previous authors have used logarithmic comparisons to account for percentage volume differences in structures that are of significantly different size (32).
Whereas the hippocampus is grossly relatively well-defined on MR images, enabling a surface reconstruction using HDM-LD, different regions of the hippocampus along the septotemporal axis are composed of distinctly different subfields (21). For example, near the septal pole, only the dentate gyrus and the CA1-3 subdivisions of the hippocampus are present. Moving from the septal pole approximately 15% of the way back toward the temporal pole, the subiculum appears. Therefore, different subregions of the hippocampus are represented uniquely along segments of the hippocampal surface. Investigators have established mutimodal, multidimensional animal brain atlases to study these complex relationships (33). By creating mathematically defined, high resolution surfaces, HDM-LD segmentations will also offer possible avenues for studying relationships of surface anatomical changes correlating to hippocampal subfields. Changes in hippocampal surfaces can be compared between groups, or longitudinally over time within subjects or groups of subjects.
We selected acquisition parameters for the rat MRI studies, including the TE = 67.5 ms, which would maximize contrast within the image, with a minimal acquisition period. Whereas past HDM-LD segmentation protocols have used T1-weighted images (14, 15), any image sequence that produces high-resolution volumetric images and provides good contrast of structures will theoretically work for HDM-LD segmentation. The current study shows that images with relative T2 weighting can also yield good HDM-LD segmentations.
Based on the methods of previous validation studies (15, 16), and the difficulties in accurate manual hippocampal segmentation of high-resolution, interpolated datasets (which typically take 2–3 h per hippocampus in our experience), we chose to limit the number of MRI datasets for our analysis. There is some ambiguity in defining borders of neuroanatomical structures on MRI (i.e., as discussed above where the hippocampus extends ventrocaudally into the temporal lobe). Therefore, interpretation of the definitive accuracy of MRI-based segmentations is difficult. Because of these factors, we chose our current methods of segmentation overlap and comparison of segmentation methods.
In conclusion, the study validates the accuracy and reproducibility of HDM-LD segmentations of the hippocampus, showing that overlap and absolute percentage volume differences between the automated HDM-LD segmentations and high-resolution manual segmentations are comparable. HDM-LD will be a useful tool for investigating in vivo hippocampal structural changes in rat models of human disease.
Funding for this project was provided in part by the Victorian Government Transport Accident Commission (TAC) in the form of a Victorian Neurotrauma Initiative grant.