| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Technical Developments |
1 From the Division of Imaging Science and Biomedical Engineering, Medical School, University of Manchester, Oxford Rd, Manchester M13 9PT, England (N.A.T., S.S., A.J.); and Cognitive Function Unit, Central Manchester Healthcare Trust, Manchester, England (A.R.V., D.B., J.S.S., D.N.). Received February 5, 2001; revision requested March 26; revision received September 17; accepted January 7, 2002. Address correspondence to N.A.T. (e-mail: neil.thacker@man.ac.uk).
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMaterials and Methods Results DiscussionAPPENDIX REFERENCES |
|---|
© RSNA, 2002
Index terms: Brain, atrophy, 13.83 • Brain, MR, 10.121411, 10.121412, 10.121419 • Dementia, 13.83 • Magnetic resonance (MR), image processing, 10.12144, 10.12149 • Magnetic resonance (MR), three-dimensional, 10.121419
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods Results DiscussionAPPENDIX REFERENCES |
|---|
In this article, we present a technique for the assessment of cerebral atrophy with magnetic resonance (MR) imaging that was designed to meet a series of criteria to enable its use in a busy clinical environment. These criteria include that the technique should provide a generic approach for the assessment of both the distribution and severity of atrophy in any brain disorder, should provide diagnostically valuable data, should be sufficiently reproducible to allow monitoring of the progression of atrophy, and should be fully automatable. To meet these design objectives, we chose to define the measurement boundaries by using a coordinate system derived from the limits of the intracranial cerebrospinal fluid (CSF) space rather than by identifying anatomic landmarks within the brain. In some ways, this approach is beneficial since the characteristics of CSF on MR images are highly distinct and can be easily identified with automatic techniques in any disease state. In addition, the CSF space reflects the premorbid head size and allows automatic correction of individual variability due to head size. However, use of the intracranial space to define the boundaries for volumetric assessment also has a number of potential disadvantages. The most obvious of these is that measurement areas within the coordinate space are not firmly related to the position of specific brain structures; therefore, the measured pattern and severity of atrophy may reflect loss of brain substance or movement of the brain within the cranium.
The purpose of this study was to establish the feasibility of distinguishing healthy subjects and patients with three dementing disorders—Alzheimer disease (AD), frontotemporal dementia (FTD), and vascular dementia (VAD)—on the basis of CSF volumetric measurements.
|
Materials and Methods |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods Results DiscussionAPPENDIX REFERENCES |
|---|
Patients with FTD or AD fulfilled currently accepted clinical diagnostic criteria for those conditions (9–11) and were free of significant risk factors for cerebrovascular disease (Hachinski scale scores, <4) (12). All patients with vascular dementia fulfilled the criteria (13) for possible vascular dementia and had high risk factors for vascular disease (Hachinski scale scores, >7). Patients exhibited the characteristic pattern of dementia associated with their clinical diagnosis (14,15). Individuals were excluded from the study if diagnosis of the form of dementia was equivocal or if the clinical pattern suggested mixed causes. None of the healthy control subjects had vascular risk factors, neurologic disease, and cognitive or psychiatric problems. Patients in the study represent a consecutive cohort of patients presenting to the clinic who fulfilled the selection criteria. The local ethics committee approved the research, and informed consent was given for inclusion into the study by the subject or legal guardian in all cases.
MR Imaging Examinations
All subjects underwent MR imaging with a 1.5-T system (ACS-NT, with PowerTrack 6000 gradient subsystem; Phillips Medical Systems, Hamburg, Germany), with a birdcage head coil receiver. CSF segmentation was performed on coronal fast spin-echo inversion-recovery images (repetition time, 6,850 msec; echo time, 18 msec; inversion time, 300 msec; echo train length, nine). Contiguous 3-mm-thick sections were obtained throughout the brain, with an in-plane resolution of 0.89 mm2 (matrix, 256 x 204; field of view, 230 x 184 mm). All imaging was performed by one of two experienced academic radiologists (A.J., A.R.V.).
Image Processing
Two researchers (N.A.T., D.B.) duplicated all manual stages of processing to investigate the effects of interobserver variability and to eliminate gross errors in data management.
Images were analyzed with machine vision software (16) to automatically identify the CSF spaces on the images. This was accomplished with derivative-based techniques to extract the edges between the CSF and other tissues and to set a threshold at the mean gray level value of the corresponding edge pixels (17). The image data were then binarized, with CSF labeled with a 1 and non-CSF tissue labeled with a 0. The binary data were then edited manually to delete tissue in the eyes and sinuses and leave CSF as the only structure within the data volume.
To place the data in a standard coordinate space, the original image volume was manually rotated. The axes of symmetry of the brain in the coronal and transverse views were aligned to the vertical and horizontal planes. The midsagittal section was then rotated to align the horizontal axis with the ventral surface of the genu of the corpus callosum and to align the junction of the corpus callosum with the body of the fornix. The resulting rotational parameters were then used to rotate the binarized CSF volume into the standardized coordinate space.
Each rotated CSF data set was used to define a rectangular CSF coordinate space that conformed to the individual head size. The anterior, posterior, lateral, and superior bounds of the CSF space were automatically identified by locating the extremes of the CSF. The inferior boundary was defined by drawing a line in the midsagittal section parallel to the horizontal axis that passed through the junction of the calvarium and the tentorium cerebelli. The resulting CSF coordinate space was divided into 12 equally sized rectangular measurement volumes that consisted of superior and inferior halves; left and right halves; and anterior, central, and posterior thirds (Fig 1). The first two planes were placed centrally along the superior-inferior and right-left axes. The third and fourth planes were coronal, and they were equidistant between the anterior and posterior boundaries of the coordinate space.
|
Data from healthy individuals were used to derive a correction for age-related atrophy by means of a linear regression to calculate the expected degree of atrophy (V): V = (age - C)/K, where C is the age at which atrophy commences and K is the rate of progression. Thus, we can construct a new variable (V') that represents the proportion of atrophy relative to that expected at a particular age: V' = VK/(age - C). By taking the value of K to be (65 - C), we can center this correction process around the age of 65 years (approximately the mean of our data set) while at the same time leaving a single free parameter with which to model the process. Values of C were taken that minimized the variance in the corrected variable V' for the healthy group.
Although correlations could be sought between individual parameters and diagnostic factors, a set of 12 measurements would require a considerable quantity of data to avoid errors due to type 2 effects and to independent correlations between the variables. Therefore, the original 12 normalized variables were corrected to remove age-related effects due to normal atrophy; they were then used to derive a set of five independent variables, each with uniform variance, before relationships with diagnostic data were examined.
The five derived variables (W1–W5) represent the age-corrected relative degree of atrophy between the central and anterior thirds of the CSF space (W1), central and posterior thirds (W2), left and right sides (W4), and dorsal and inferior halves (W5). The remaining variable (W3) is a measure of the overall relative degree of atrophy. These variables represent a rotation in the original space that was chosen to allow correlations between the variables to be effectively eliminated in all but W3. The methods of derivation of these five derived variables and the rationale behind the data reduction steps are described in the Appendix.
The five derived variables (W1–W5) were then used to construct a Parzen window classifier by using the repeatability estimates from the reproducibility study to scale the probability kernel. The performance of this classifier was evaluated with a leave-one-out cross-validation technique. The distribution of the variables was also visualized by using a software tool (18) to examine the data for strong correlations or diagnostic patterns and to produce illustrations of the identified diagnostic trends.
|
Results |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods Results DiscussionAPPENDIX REFERENCES |
|---|
|
|
|
|
The last variable, W5, shows no obvious discrimination capabilities when used in conjunction with other parameters, although the distribution of this parameter between groups still holds some residual information. Patients with AD have more specific atrophy in the superior half of the coordinate space than do the other groups, and patients with VAD have relatively more atrophy in the inferior compared with the superior half of the coordinate space.
Diagnostic Performance of the Normalized Variables
To obtain a quantitative measure of separability of the data, the repeatability measures from Table 2 were used to define the kernel for a Parzen classifier. Defined this way, the classifier is computing the relative likelihood that each pattern can be considered to be consistent with each data class on the basis of measurement accuracy. In this analysis, each subject was entered twice owing to the pooling of data from the repeatability study.
Table 3 shows the first set of diagnoses, with each data point excluded from each subject as it was classified, which represent a lower limit on classification performance. The second set of diagnoses, with exclusion of only the measurements being classified, are a fair reflection of performance on the basis of the assumption that two observations of the same subject are independent. This assumption would be true if the SDs and repeatabilities were equivalent; given their interpretation as upper and lower limits on accuracy, we believe that this assumption is reasonable. Table 3 is useful for assessment of the distribution and separability of our patient groups in the five-dimensional classification space; however, the main diagnostic problems in clinical practice are the separation of AD and FTD and the diagnosis of early AD in patients compared with normal findings in healthy elderly subjects.
|
|
|
Discussion |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods Results DiscussionAPPENDIX REFERENCES |
|---|
A more logical approach to quantification of atrophy is measurement of the volumes of cerebral structures, which requires delineation of structure margins. This delineation is commonly accomplished with manual tracing techniques, which are extremely time-consuming. Manual segmentation of the hippocampus on a coronal data set with contiguous 1.5-mm-thick sections takes approximately 10–15 minutes depending on the expertise of the operator. Nonetheless, it is possible to train observers to attain a high degree of reproducibility, and this approach remains the method of choice for volumetric assessment of well-delineated structures. This approach limits the number of structures that can be studied, however, and forces an a priori decision about the structures that will be included in the analysis.
In patients with AD with atrophy of medial temporal lobe structures, particularly the hippocampus, entorhinal cortex and uncus occur early in the disease and progress as much as 10 times faster than is seen in healthy elderly subjects (30–33). Measurements of the hippocampal and entorhinal cortex volumes distinguish patients with AD from healthy individuals and from patients with depression or other neurodegenerative brain diseases, with a specificity of greater than 95% (19,34,35). Unfortunately, medial temporal lobe atrophy is also a feature of other neurodegenerative diseases (4,20). In FTD, the brain tends to be more atrophic than in AD (21,36), and medial temporal lobe atrophy is prominent, although it is proportionately less severe than that seen in AD (20,37). In one study, measurement of entorhinal cortex and hippocampal volumes resulted in diagnostic sensitivities of less than 60% for FTD and 80% for AD (20).
To avoid dependence on measurements of single structures, development of a metric that reflects the overall pattern and severity of cerebral atrophy and is applicable in any disease state is desirable. Automatic identification of brain structures is possible with use of the concepts of anatomic trained models (38). However, trained models cannot cope with variation outside the initial training set of data. In practice, an extensive training set that included examples of all atrophy patterns that might be identified in the patient population would need to be developed.
An alternative approach to providing a generic metric for the assessment of whole-brain atrophy is identification of structures by transforming the image data into a standard sampling space, as described by Talairach and Tournoux (39), in which the spatial distribution of the structure within the space is known. This approach has been used to automate the measurement of structural volumes in healthy subjects and in patients with schizophrenia (40). However, this approach does not take into account any spatial distortion induced by the atrophic process and is applicable only in disorders where atrophic changes are relatively small and are not associated with any known spatial distortion.
Our approach is deliberately simplistic and generic. We identified a three-dimensional coordinate space based on the size and shape of the cranial cavity that was identified by means of segmentation of CSF from the data set. Use of this coordinate system to define sampling volumes removes any sampling bias and provides automatic normalization for individual variations in head volume. Use of CSF as the marker tissue allows the use of simple and robust segmentation techniques that will not be affected by disease-related changes in gray and white matter properties. Use of the intercranial space as a basis for the coordinate system makes the technique highly generic and free from systematic errors that might result if spatial distortion of the brain were to result from specific disease processes. The technique is capable of full automation, can automatically identify the most likely diagnosis, and can assign a probability to that diagnosis that can be used to guide the clinician or researcher in the clinical diagnostic process.
This technique is highly reproducible, and the effect of measurement errors is seen in Table 2. These errors represent upper limits on actual measurement performance. The reproducibility estimate (SD [Wi - Wi']) gives an indication of the ability to monitor the progression of atrophy. Data that differ by more than these quantities can be said to be measurably different. The similarity between the scales of repeatability and SDs of the healthy group is striking. In the healthy group, intergroup variations for approximately half can be accounted for by measurement error. This factor might be reduced with full automation of all stages of data processing.
The technique showed good diagnostic performance in the patient groups that we studied, particularly when specific diagnostic questions were addressed. The technique helped distinguish AD patients from a mixed population of AD patients and healthy age-matched control subjects with a sensitivity of 95% and specificity of 93%; these findings are equal to or better than the best results reported on the basis of medial temporal lobe volume measurements from computed tomography or MR imaging (20,41–43). This performance was improved by applying a Bayesian probability threshold of 80% to produce sensitivities and specificities of 97%. When these diagnostic performances are considered, it is important to note that the AD patients in this study were young (age range, 49–73 years; median age, 61 years) and many had early disease (time from onset, 2–8 years; median time from onset, 3 years). The FTD syndrome is less well recognized than AD; although the criteria for diagnosis have been well established (10,11), they require additional assessment by a neuropsychologic specialist, which may be impractical in nonspecialist centers. The technique described here helped identification of FTD patients from a population of AD or FTD patients with a sensitivity of 79% and specificity of 72%, which increased to 82% and 92%, respectively, after application of an 80% probability threshold. These performance data support the use of this approach as a diagnostic tool to help identify demented patients in whom a diagnosis of AD should be reviewed.
Although the results of this study were clinically promising, they must be considered proof of concept to support the use of a generic approach to quantification of cerebral atrophy assessment as a potential clinical tool. In this study, the rotational coordinates of the data space were manually defined from structures identified within the brain. In clinical practice, we would aim to automate this transformation step by means of automated alignment of the intercranial cavity itself into the data space by using surface-matching or trained-model–based techniques. This automation would remove the requirement for manual editing of extracranial fluid structures and would also be likely to reduce measurement errors. The segmentation technique we used is simplistic, does not account for partial volume-averaging effects, and would be inappropriate for the segmentation of gray and white matter. Several of the parameters that we used were chosen arbitrarily. The size of the rectangular coordinate space was used to correct for individual head size, whereas the actual volume of the intercranial space may provide a more robust normalization factor. The divisions of the data space were arbitrarily selected to provide equal sampling volumes that covered the entire prosencephalon. In practice, the existence of different disease-specific patterns of atrophy suggests that modification of sample selection would improve results in a wide range of atrophic patterns.
One possible improvement would be the inclusion of a central sample that would specifically reflect ventricular involvement. The derivation of optimal sampling strategies would be disease specific and could be best established by using genetic algorithms, which require larger data sets than were available for the present study. Finally, the age correction that we applied is based on measurements from only nine cases and requires validation with a much larger population. Other potential improvements in discrimination include segmentation of brain tissue to help identify the relative severity and distribution of loss of gray and white matter that contributes to the overall atrophic pattern, which varies in different disease states (21,37,44,45). Finally, the use of a Bayesian probabilistic approach that was based on findings in a sample of cases that were selected to represent definite diagnoses imposes restrictions on the interpretation of the diagnostic value of the technique. The diagnostic performance documented in this study represents the ability to differentiate between cases in which detailed clinical and neuropsychologic assessment alone would have provided a confident diagnosis. This results from the requirement to use clinical diagnosis as a standard of reference in the absence of brain biopsy or postmortem evidence. This restriction also applies to other imaging-based diagnostic studies because biopsy or follow-up after death are rarely available. Diagnostic performance in very early cases or in more advanced cases in which the clinical diagnosis was uncertain cannot be assessed with these data. Therefore, correct training of this type of classifier will require the addition of cases of this type (ie, confirmed at further follow-up or at pathologic examination) to the training set.
If this approach to diagnostic decision support is to be used in clinical practice, a number of other steps are essential. The results of this study must be tested with other larger clinical groups. The atrophy patterns of a number of other clinical conditions that might cause diagnostic confusion, such as normal pressure hydrocephalus and Lewy body dementia, must also be examined. The sensitivity of the technique to longitudinal atrophic changes should be established, and the technique itself must be fully automated, optimized, and revalidated to allow routine use. If these preliminary results are supported, then this method provides a new diagnostic decision support system that is practical in even a busy clinical environment. The method we describe will not provide a stand-alone clinical tool; instead, it would be used in combination with existing clinical techniques. Other diagnostic imaging techniques, particularly studies of cerebral perfusion patterns, may provide additional uncorrelated diagnostic information that can be used to increase the probability of each diagnostic classification.
In conclusion, we have demonstrated that automated methods to describe the severity and distribution of cerebral atrophy are capable of providing diagnostic information in the classification of neurodegenerative diseases. The method we developed is intended only as a proof of concept and requires considerable refinement and validation before it can be used in clinical practice. However, the technique has been designed to be easily automatable and could realistically form the basis of a practical diagnostic decision support system.
|
APPENDIX |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods Results DiscussionAPPENDIX REFERENCES |
|---|
Calculation of Final Derived Variables
In repeatability studies, the age-corrected variables were found to have the approximate characteristics of Poisson random variables (ie, the data had increasing variance for larger measurements). Thus, these measures can be converted into measures that have uniform variance by taking the square root of the value (46). The set of seven square root variables should define a seven-dimensional space with homogeneous errors. However, residual intrasubject age differences, developmental effects, and systematic measurement errors leave some residual correlations between these absolute factors. Therefore, a final set of five derived variables (W1–W5) were calculated from the intermediate variables (A, P, C, R, L, S, and I). These final derived variables are corrected for head size variation and for healthy age-related cerebral atrophy, have no residual independent correlations (with the exception of W3), and have uniform variance. These derived variables were calculated as follows:
1. The age-corrected relative degree of atrophy between the central and anterior thirds of the coordinate space (W1):
2. The age-corrected relative degree of atrophy between the central and posterior thirds of the coordinate space (W2):
3. The age-corrected relative degree of total atrophy (W3):
4. The age-corrected relative degree of atrophy between the left and right halves of the coordinate space (W4):
5. The age-corrected relative degree of atrophy between the superior and inferior halves of the coordinate space (W5):
|
FOOTNOTES |
|---|
Author contributions: Guarantors of integrity of entire study, A.J., D.N.; study concepts, A.J., N.A.T., A.R.V.; study design, A.J., N.A.T., D.N.; literature research, A.J., A.R.V., N.A.T., S.S.; clinical studies, A.R.V., D.N., J.S.S.; data acquisition, A.J., A.R.V., J.S.S.; data analysis/interpretation, N.A.T., D.B., A.R.V., S.S.; statistical analysis, N.A.T.; manuscript preparation, N.A.T., A.J.; manuscript definition of intellectual content, N.A.T., A.R.V., A.J., J.S.S.; manuscript editing, A.J.; manuscript revision/review, all authors; manuscript final version approval, N.A.T.
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods Results DiscussionAPPENDIX REFERENCES |
|---|
This article has been cited by other articles:
|
|
 |
|
  S MILLS, J CAIN, N PURANDARE, and A JACKSON Biomarkers of cerebrovascular disease in dementia Br. J. Radiol., December 1, 2007; 80(Special_Issue_2): S128 - S145. [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]
|
|
|
|
|
|
T. F. Patankar, D. Mitra, A. Varma, J. Snowden, D. Neary, and A. Jackson Dilatation of the Virchow-Robin Space Is a Sensitive Indicator of Cerebral Microvascular Disease: Study in Elderly Patients with Dementia AJNR Am. J. Neuroradiol., June 1, 2005; 26(6): 1512 - 1520. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 R. BALDWIN, S. JEFFRIES, A. JACKSON, C. SUTCLIFFE, N. THACKER, M. SCOTT, and A. BURNS Neurological findings in late-onset depressive disorder: comparison of individuals with and without depression The British Journal of Psychiatry, April 1, 2005; 186(4): 308 - 313. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| RADIOLOGY | RADIOGRAPHICS | RSNA JOURNALS ONLINE |
). More extensive atrophy is seen in patients with AD (
) and FTD (+). Also note the separate distribution of patients with FTD or AD, which indicates distinct patterns of atrophy in each disease. (b) Plot of W1 (relative degree of atrophy between the anterior and central thirds of the coordinate space) versus W4 (left-right asymmetry). Note the presence of asymmetry in many patients with AD (