| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Technical Developments |
1 From the Department of Neurology (R.E.H.) and the Division of Neurosurgery (R.D.B.), Saint Louis University, 3635 Vista Ave, St Louis, MO 63110; IntellX, Broomfield, Colo (K.E.M., S.J., R.D.B.); the Department of Psychiatry, Washington University School of Medicine, St Louis, Mo (L.W.); and the Center for Imaging Science, Whiting School of Engineering, Johns Hopkins University, Baltimore, Md (M.I.M.). Received May 17, 1999; revision requested July 19; final revision received October 12; accepted October 26. Address correspondence to R.E.H. (e-mail: hoganr2@slu.edu).
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMaterials and MethodsResultsDiscussionREFERENCES |
|---|
Index terms: Brain, volume, 1341.839 • Epilepsy • Hippocampus, 1341.92 • Magnetic resonance (MR), image processing, 1341.121412 • Magnetic resonance (MR), volume measurement, 1341.12141 • Schizophrenia
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and MethodsResultsDiscussionREFERENCES |
|---|
Past studies have used manual segmentation of the hippocampus on MR images to determine hippocampal volumes (8). Although the sensitivity for detecting hippocampal asymmetry of manual segmentation as compared with visual inspection is greater in a proportion of patients with mesial temporal lobe epilepsy (9,10), 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 anatomic variations. The emerging field of computational anatomy founded on general pattern theory (11) provides tools and a framework for accommodating and studying this variability (12–14). In this framework, an electronic atlas of the hippocampus is used as a deformable template that is matched to an individual MR image to extract and study the individual hippocampal areas.
Haller et al (15,16) describe a deformation-based hippocampal segmentation technique and verify the precision of this technique in healthy and schizophrenic patients. In this study, we evaluated the precision and reproducibility of deformation-based hippocampal segmentations in a group of patients with confirmed mesial temporal lobe epilepsy and mesial temporal sclerosis, which result in noteworthy changes in the size and shape (17) of the hippocampus.
|
Materials and Methods |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and MethodsResultsDiscussionREFERENCES |
|---|
Five patients (three men and two women) with refractory epilepsy were evaluated with MR imaging. Epilepsy in all patients had been refractory to multiple pharmacologic treatments. Mean duration of epilepsy at the time of epilepsy surgery was 9.4 years (range, 4.6–15.4 years). Mean age at the time of epilepsy surgery was 32.2 years (range, 25–39 years). All patients were seizure free after epilepsy surgery (mean postoperative follow-up, 23.7 months; range, 18–33 months), and mesial temporal sclerosis was confirmed at postsurgical pathologic examination.
For manual hippocampal segmentation, MR images were integrated into an independent computer workstation and analyzed with software (ANALYZE AVW, version 1.1; Biomedical Imaging Resource, Mayo Foundation, Rochester, Minn). MR images were converted to an isotropic voxel dimension of 0.859. Intensities on these images were then adjusted to match those on the atlas image. For each image, a region of interest (40 x 40 x 64) was outlined for each hippocampus. This region of interest was converted to a voxel size of 0.4295 by means of interpolation, and final dimensions were 80 x 80 x 128.
Two investigators (R.E.H., L.W.) independently segmented the hippocampi. Figure 1 shows general orientation of the MR image with respect to the outline of the region of interest. The hippocampi were segmented on the basis of anatomic boundaries described by Watson et al (18), with some exceptions. Watson et al used only coronal sections in their hippocampal tracings. We examined and segmented the hippocampus in coronal, sagittal, and transverse planes; we reviewed segmentations in each plane twice to ensure accurate tracings in all three dimensions. In Tables 1 and 2, patients' hippocampal images were referred to as 1R (patient 1, right hippocampal image), 1L (patient 1, left hippocampal image), and so on.
|
|
|
|
Verification of hippocampal head regions in the sagittal and transverse planes was often useful. In the sagittal plane, the medial aspect of the uncinate gyrus, the most medial part of the hippocampal head, was identified in the plane where the medial margin of the infolding of the uncal cleft was present inferiorly, and the semilunar gyrus was present superiorly (19). The semilunar gyrus often was present as a protuberance in the superior aspect of the amygdalohippocampal complex in this plane. Progressing from this plane laterally, the outline of the anterosuperior boundary of the hippocampal head was again verified in relationship to the uncal recess of the lateral ventricle and the alveus. The uncal cleft and subiculum were included in the inferior region of the segmentations. The transverse plane was often useful in defining the anterolateral region of the hippocampal head in relation to the lateral ventricle. Figure 2 is an image where the alveus is visible in the region of the hippocampal head. Figure 3 illustrates the use of landmarks to help define the hippocampal head when the alveus or the uncal recess of the lateral ventricle is not visible. Time for manual segmentation was approximately 2 hours per hippocampus.
|
|
|
Determination of landmarks provided an initial condition for the intensity-matching algorithm by roughly aligning the patient and atlas images. The first step in determining landmarks was identification of global landmarks, which scale and align the atlas brain to the patient brain on the basis of standard landmarks from the coordinate system of Talairach and Tournoux (20). The second step was individual determination of landmarks on each hippocampus. This was done by first identifying the head and tail of the hippocampus, which specifies the long axis of the hippocampus. Then, four landmarks were identified on five cross sections equally spaced along this axis. Landmarks were placed on the medial, lateral, superior, and inferior borders of the hippocampus on each cross section.
Images and landmark data were then integrated into a Unix-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 image to a patient MR image. The atlas reference MR image was segmented as previously described (16). The coarse aspect of the procedure relied on the landmark information provided by expert segmenters to provide an initial low-dimensional coregistration of atlas and patient images (21). The landmark information was provided in the form of the global and hippocampus-specific landmarks, which were used to derive a coarse manifold transformation (22) from the reference to the patient images.
After the coarse first step in the transformation was completed, the volumes were roughly aligned and attention was focused on the fine-featured substructures. The fine procedure involved the next two steps. The second step was to solve the registration problem by means of a linear elastic basis formulation and the full-volume data, as previously described (13,23). This was fully automatic and driven by only the volume data itself. The three-dimensional whole-brain maps corresponded to the maximizer, whose variation solution corresponded to a solution of a nonlinear partial differential equation, consisting of between 107 and 108 parameters. The third and final step of the algorithm was to solve the nonlinear partial differential equation corresponding to the Bayesian maximizer associated with the fluid formulation at each voxel of the full volume (12,24,25).
For the deformation-based segmentations, each data set was completely preprocessed (adjustment of intensity and determination of landmarks) twice at an interval of 2 months and run through the deformation-based algorithm.
Comparisons between two segmentations were made by computing the percentage overlap of voxels (Figs 4, 5). One segmentation was designated the reference (R) and the other the study (S) segmentation that we compared against the reference. The percentage 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 x 100. We used the manual segmentations as the reference segmentations. Because a relatively larger or smaller reference image would effect the percentage overlap, automated segmentations were compared separately with the manual segmentations from both investigators. Between-group comparison of volume estimates was made by means of a paired Student t test (one tailed).
|
|
|
Results |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and MethodsResultsDiscussionREFERENCES |
|---|
The means and SDs of the percentage overlap were computed for three subsets of the data: normal hippocampi, sclerotic hippocampi, and all hippocampi. Overall comparisons showed a mean percentage overlap of 92.8% (SD, 3.5) between A1 and A2, 73.1% (SD, 9.5) between M1 and M2, and 74.8% (SD, 10.3) between A1 and M1. The relatively large percentage overlap between A1 and A2 demonstrates the reproducibility of the deformation-based segmentation procedure. The last two vales demonstrate that the automatic segmentation procedure performed comparably with manual segmentation.
The mean percentage overlap for the subset of normal hippocampi were 80.8% (SD, 4.6) between M1 and M2 and 82.9% (SD, 4.4) between A1 and M1. The mean percentage overlap for the subset of sclerotic hippocampi were 65.4% (SD, 5.9) between M1 and M2 and 66.6% (SD, 7.4) between A1 and M1. The decrease in mean percentage overlap between segmentation of normal versus sclerotic hippocampi reflects the increased ambiguity and difficulty in segmenting the sclerotic hippocampi.
Table 2 shows volume measurements based on the four segmentations. Mean values for absolute percentage differences for normal hippocampi were the following: A1 versus A2, 4.5% (SD, 1.9%); M1 versus M2, 10.4% (SD, 8.0%); A1 versus M1, 6.2% (SD, 1.7%). Mean values for absolute percentage differences for sclerotic hippocampi were the following: A1 versus A2, 4.0% (SD, 3.5%); M1 versus M2, 16.2% (SD, 13.7%); A1 versus M1, 15.8% (SD, 5.7%). Mean values for absolute percentage differences for all hippocampi were the following: A1 versus A2, 4.3% (SD, 2.7%); M1 versus M2, 13.3% (SD, 11.0%); A1 versus M1, 11.0% (SD, 6.4%). Results of volume comparisons also showed greater differences in the sclerotic hippocampi.
Finally, comparisons of percentage overlap, considering all hippocampi, were calculated with use of either M1 or M2 as the reference. The comparisons to M1 as the reference showed that the automatic segmentations had a mean percentage overlap of 74.8% (M1 vs A1) and 78.5% (M1 vs A2) as compared with 73.1% (M1 vs M2). Similarly, with M2 as the reference, the mean percentage overlap of A1 (M2 vs A1) was 74.9% and the of A2 (M2 vs A2) was 76.9%.
|
Discussion |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and MethodsResultsDiscussionREFERENCES |
|---|
Past techniques measuring the hippocampus have focused on defining structures in a single two-dimensional plane. There are advantages of measuring the hippocampus in the coronal plane, which is perpendicular to the long axis of the hippocampal body (18). Certain hippocampal structures, such as the boundary between the subiculum and parahippocampal gyrus in the body of the hippocampus, are most easily identifiable in the coronal plane. Because of the three-dimensional shape of the hippocampus, however, especially the hippocampal head, which turns medially and superiorly in relationship to the hippocampal body, we have found advantages in verifying hippocampal segmentations in coronal, sagittal, and transverse planes. Because the boundaries of the hippocampal head were the most difficult to define, verification of three-dimensional relationships in this region was most helpful. The indistinct boundary between the uncinate gyrus of the hippocampus and ambient gyrus of the amygdala was often best determined by using landmarks in the sagittal plane, as described in Materials and Methods. Figure 3 illustrates this region. Other investigators have commented on the advantages of segmentation of the hippocampal head region by means of three-dimensional segmentations (27,28).
Our results show reproducibility and are comparable to those in past studies. Direct comparison with past validation studies is difficult, however, owing to differences in acquisition parameters of images in different studies (27), past documentation of validation in only healthy subjects (29), and differences in anatomic boundaries and segmentation techniques used for segmentations (8). These differences have produced a wide variation in normal hippocampal volumes in different studies (30).
Most past validation studies have used manual segmentation techniques (27–29). With use of general pattern matching, anatomic structures can be segmented on the basis of global shape models. By using templates that represent the typical structures, MR images of the hippocampus may be semiautomatically segmented by means of template transformations to define hippocampal variability (12,13,22). Haller et al (15) reported a method of hippocampal segmentation based on general pattern matching. In a comparison of two-dimensional measurements of a segment of the hippocampus on the basis of general pattern matching, the mean difference between two segmentations was 1.33%, while that between two manual segmentations was 4.67%.
Our validation is most comparable to that performed by Haller et al (16), in which investigators validated a deformation hippocampal segmentation technique by comparing control subjects and patients with schizophrenia. They performed two deformation and two manual segmentations on MR images of the right hippocampus in five healthy and five schizophrenic patients. The mean percentage difference between absolute volumes for automatic and manual segmentations, respectively, were 3.6% and 4.2% for healthy subjects and 2.5% and 10.1% for schizophrenic patients. Between-method mean percentage differences in volumes (with use of nonabsolute values) for automated versus manual segmentations were -0.7% (range, 6.4% to -12.2%; SD, 7.0%) for healthy subjects and -2.4% (range, 14.1% to -11.1%; SD, 10.6%) for schizophrenic patients. Overall percentage overlap (considering both healthy and schizophrenic groups together) between automated segmentations was 91.4% (SD, 3.6%), between manual segmentations was 77.9% (SD, 4.8%), and between automated and manual segmentations was 74.2% (SD, 5.5%). The between-method percentage differences for absolute volumes and overall percentage overlap were calculated in a similar manner in our study and the study of Haller et al (16). Our results for differences in absolute percentage volume and percentage overlap were comparable. In a comparison of automated versus manual segmentations, our between-method absolute percentage differences in volumes were 6.2% (range, 4.4%–8.6%; SD, 1.7%) for normal hippocampi and 15.8% (range, 5.7%–18.6%; SD, 5.7%) for sclerotic hippocampi. Our values for overall percentage overlap between automated segmentations were 92.8% (SD, 3.5%), between manual segmentations was 73.1% (SD, 9.5%), and between automated and manual segmentations was 74.8% (SD, 10.3%). Differences in percentage overlap and percentage volume were similar between the manual segmentations and the automated versus manual segmentations, demonstrating a similar variability in the automatic and manual segmentations when compared with the same control.
In our study, as in the study by Haller et al (16), results in normal hippocampi were in better concordance than were those in abnormal hippocampi with both automated and manual segmentations. Past studies have shown greater percentage error when segmentations involve smaller volumes (31). In a study of progressive multiple sclerosis lesions depicted at MR imaging, Goodkin et al (31) found a coefficient of variation for three successive lesion measurements that was inversely related to the lesion area and ranged from 22.6% for lesions smaller than 0.67 cm2 to 12.1% for larger lesions. Past authors have used logarithmic comparisons to account for percentage volume differences in structures that are of different size (28). Therefore, greater variance could be expected in the segmentation of sclerotic hippocampi because they are smaller than normal hippocampi.
Another possible source of error is that deformation mapping of the normal hippocampus onto an atrophic hippocampus may be slightly more difficult for the mapping algorithm than is mapping onto a normal hippocampus. This could possibly increase variability in deformation segmentations of sclerotic hippocampi. Determination of landmarks is also variable in automated segmentations and is also more difficult in atrophic hippocampi. Our results indicate increased difficulty in defining hippocampal borders with either the deformation or manual segmentation technique when the hippocampus is small or abnormally shaped.
A comparison of percentage overlap of voxels with use of either M1 or M2 as the reference did not show appreciable differences. Given the formula for calculating percentage overlap, a relatively larger reference would produce a higher percentage overlap, and a relatively smaller reference would produce a relatively lower percentage overlap. The comparable values of percentage overlap calculated with either M1 or M2 as a reference indicate that the size of M1 and M2 is comparable. This was confirmed in the comparisons of volume.
There are many clinical applications for MR-based hippocampal volumetric studies. Many studies continue to explore applications in the areas of epilepsy, Alzheimer disease (32,33), amnesic syndromes (34), and schizophrenia (35,36). We demonstrated that our deformation-based segmentation technique produces reliable segmentation of the hippocampi in patients with mesial temporal sclerosis and mesial temporal lobe epilepsy. Practically, this semiautomated technique allows segmentation of both hippocampi in approximately 10–15 minutes of user time, which is shorter than the time needed to produce accurate segmentations with manual segmentation—approximately 2 hours per hippocampus in this study. Other investigators performing segmentation with images of different voxel sizes report taking approximately 40 minutes per hippocampus (27).
Deformation-based segmentation can help examine three-dimensional aspects of hippocampal shape. Analysis of changes in shape of schizophrenic versus normal hippocampi showed changes, but volume changes alone were not different (21,35,37). Analysis of hippocampal shape may also have important applications in the diagnosis of other central nervous system disorders. Shape analysis enables evaluation of local details of hippocampal anatomy that may not be evident in measurements of total hippocampal volume (Fig 6). In epilepsy, hippocampal sclerosis may be present without associated asymmetry in total hippocampal volume (38). There are also well-documented changes in regional hippocampal anatomy, such as volume loss localized to only the hippocampal head (17,39) or loss of digitations of the hippocampal head (40). For the latter in MR diagnosis of mesial temporal sclerosis, sensitivity was 92% and specificity was 100%. Figure 7 illustrates the way deformation-based segmentations can highlight structural changes in the hippocampal head. Application of deformation shape analysis may enable better localization of neuroanatomic abnormalities in patients with mesial temporal lobe epilepsy.
|
|
|
ACKNOWLEDGMENTS |
|---|
|
FOOTNOTES |
|---|
Author contributions: Guarantor of integrity of entire study, R.E.H.; study concepts, R.E.H.; study design, R.E.H., K.E.M.; definition of intellectual content, R.E.H., K.E.M., M.I.M., S.J.; literature research, R.E.H.; clinical studies, R.E.H., R.D.B.; data acquisition, R.E.H.; data analysis, R.E.H., K.E.M., L.W.; statistical analysis, K.E.M.; manuscript preparation, R.E.H., K.E.M., S.J.; manuscript editing, R.E.H.; manuscript review, R.E.H., K.E.M., R.D.B.
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and MethodsResultsDiscussionREFERENCES |
|---|
This article has been cited by other articles:
|
|
 |
|
  O. Colliot, G. Chetelat, M. Chupin, B. Desgranges, B. Magnin, H. Benali, B. Dubois, L. Garnero, F. Eustache, and S. Lehericy Discrimination between Alzheimer Disease, Mild Cognitive Impairment, and Normal Aging by Using Automated Segmentation of the Hippocampus Radiology, July 1, 2008; 248(1): 194 - 201. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 R E Hogan, R P Carne, C J Kilpatrick, M J Cook, A Patel, L King, and T J O'Brien Hippocampal deformation mapping in MRI negative PET positive temporal lobe epilepsy J. Neurol. Neurosurg. Psychiatry, June 1, 2008; 79(6): 636 - 640. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 B. H. Ridha, J. Barnes, L. A. van de Pol, J. M. Schott, R. G. Boyes, M. M. Siddique, M. N. Rossor, P. Scheltens, and N. C. Fox Application of Automated Medial Temporal Lobe Atrophy Scale to Alzheimer Disease Arch Neurol, June 1, 2007; 64(6): 849 - 854. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 R.E. Hogan, L. Wang, M.E. Bertrand, L.J. Willmore, R.D. Bucholz, A.S. Nassif, and J.G. Csernansky Predictive Value of Hippocampal MR Imaging-Based High-Dimensional Mapping in Mesial Temporal Epilepsy: Preliminary Findings AJNR Am. J. Neuroradiol., November 1, 2006; 27(10): 2149 - 2154. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
F.L. Giesel, H.K. Hahn, P.A. Thomann, E. Widjaja, E. Wignall, H. von Tengg-Kobligk, J. Pantel, P.D. Griffiths, H.O. Peitgen, J. Schroder, et al. Temporal Horn Index and Volume of Medial Temporal Lobe Atrophy Using a New Semiautomated Method for Rapid and Precise Assessment AJNR Am. J. Neuroradiol., August 1, 2006; 27(7): 1454 - 1458. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 R. E. Hogan, L. Wang, M. E. Bertrand, L. J. Willmore, R. D. Bucholz, A. S. Nassif, and J. G. Csernansky MRI-based high-dimensional hippocampal mapping in mesial temporal lobe epilepsy Brain, August 1, 2004; 127(8): 1731 - 1740. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
K. Kaiboriboon and R. E. Hogan Hippocampal Shape Analysis in Status Epilepticus Associated with Acute Encephalitis AJNR Am. J. Neuroradiol., June 1, 2002; 23(6): 1003 - 1006. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| RADIOLOGY | RADIOGRAPHICS | RSNA JOURNALS ONLINE |
