Radiology
HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
QUICK SEARCH:   [advanced]


     

Published online before print January 25, 2002, 10.1148/radiol.2223010413
This Article
Right arrow Abstract Freely available
Right arrow Figures Only
Right arrow Full Text (PDF)
Right arrow All Versions of this Article:
2223010413v1
222/3/722    most recent
Right arrow Submit a response
Right arrow Alert me when this article is cited
Right arrow Alert me when eLetters are posted
Right arrow Alert me if a correction is posted
Right arrow Citation Map
Services
Right arrow Email this article to a friend
Right arrow Similar articles in this journal
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Dehmeshki, J.
Right arrow Articles by Tofts, P. S.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Dehmeshki, J.
Right arrow Articles by Tofts, P. S.
(Radiology 2002;222:722-728.)
© RSNA, 2002

Neuroradiology

Systemic Lupus Erythematosus: Diagnostic Application of Magnetization Transfer Ratio Histograms in Patients with Neuropsychiatric Symptoms—Initial Results1

Jamshid Dehmeshki, MSc, PhD, Mark A. Van Buchem, MD, PhD, Gerlof P. T. Bosma, MD, Tom W. J. Huizinga, MD, PhD and Paul S. Tofts, BA, DPhil

1 From the Institute of Neurology, University College London, England (J.D., P.S.T.); and Departments of Radiology (M.A.V.B., G.P.T.B.) and Rheumatology (T.W.J.H.), Leiden University Medical Center, the Netherlands. Received February 5, 2001; revision requested March 16; final revision received September 6; accepted September 19. Address correspondence to J.D., Medicsight plc, 46 Berkeley Square, London W1J 5AT, England (e-mail: j.dehmeshki@http-tech.com).


   ABSTRACT
TOP
 ABSTRACT
INTRODUCTION
 MATERIALS AND METHODS
RESULTS
 DISCUSSION
 APPENDIX
 REFERENCES
 
PURPOSE: To explore the diagnostic potential of magnetization transfer ratio (MTR) histogram analysis in patients with neuropsychiatric systemic lupus erythematosus (SLE) by using multivariate discriminant analysis (MDA).

MATERIALS AND METHODS: Volumetric magnetization transfer imaging was performed in nine patients with active nonthromboembolic, neuropsychiatric SLE, 10 patients with SLE who had had neuropsychiatric SLE previously, 10 patients with SLE but no history of neuropsychiatric SLE, 10 patients with inactive multiple sclerosis, and 10 healthy control subjects. For each subject, an MTR histogram of the whole brain was generated, and an MDA score was produced for each histogram. Each patient was assigned to a clinical subgroup on the basis of these MDA scores. For assignment, binary comparisons between subgroups were made. The accuracy of this classification method was assessed and compared with that of conventional MTR histogram analysis.

RESULTS: With MDA, the success rate of binary classification was 60%–100%, depending on which two groups were compared. When the different clinical subgroups were separated, MDA parameters were always better than conventional MTR histogram parameters, with P values ranging from .05 to less than 1 x 10-6 of those attained with the best conventional parameter.

CONCLUSION: With MDA, MTR histograms of brain tissue may provide diagnostic information for individual patients in the clinical context of SLE.

© RSNA, 2002

Index terms: Brain, diseases, 13.61 • Brain, MR, 13.121411, 13.121412, 13.121413, 13.121416, 13.121417 • Lupus erythematosus, 13.61 • Magnetic resonance (MR), magnetization transfer contrast, 13.121417


   INTRODUCTION
TOP
 ABSTRACT
 INTRODUCTION
MATERIALS AND METHODS
RESULTS
 DISCUSSION
 APPENDIX
 REFERENCES
 
In patients with systemic lupus erythematosus (SLE), neurologic and psychiatric symptoms occur in up to 75% of cases during the course of the disease (1,2). These symptoms may be induced by drugs that are used to suppress the underlying disease process, such as steroids, and they may be secondary to SLE. As a consequence of antiphospholipid antibodies, an increased tendency for thromboembolic events may occur, and this can lead to cerebral infarctions. Furthermore, neurologic and psychiatric symptoms can occur without evidence of cerebral infarction at imaging, and they might not be attributed to medication. In such cases, an intrinsic SLE-related brain process presumably is active (3).

Because many patients with SLE are taking medication and have antiphospholipid antibodies at the same time, the occurrence of neurologic and psychiatric symptoms may create a diagnostic challenge, and a determination of the origin of the symptoms may be impossible when it is based solely on clinical information. This uncertainty of the origin of the symptoms is particularly troubling, because if intrinsic brain involvement is suspected, medications such as chemotherapeutic agents with severe side effects may be prescribed. Therefore, a misdiagnosis that results in over- or undertreatment is a serious risk for patients who have SLE with neurologic and psychiatric symptoms, or neuropsychiatric SLE (NPSLE).

Bosma and co-workers (4) recently found that differences between various clinical subgroups of patients with SLE could be detected by using magnetization transfer imaging. Magnetization transfer imaging enables assessment of the amount of transfer of magnetization between pools of bound and free protons. This exchange process can be quantified by calculating the magnetization transfer ratio (MTR), and it has been demonstrated that MTRs reflect the type and degree of histologic changes in brain tissue. By using magnetization transfer imaging, Bosma and co-workers (5) detected differences in brain tissue characteristics between a group of patients with SLE who had had neurologic and psychiatric symptoms in the past (past NPSLE group) and a group of patients with SLE who had never had neurologic and psychiatric symptoms (non-NPSLE group). They also detected differences between a group of patients with NPSLE during an episode of neurologic and psychiatric symptoms (active NPSLE group) and the following groups: patients with past NPSLE, patients with non-NPSLE, patients with multiple sclerosis (MS), and healthy control subjects (4,5).

In these studies, the MTRs of all pixels representing the brain were displayed as MTR histograms, and the histogram characteristics were compared between the groups by using the method described by Van Buchem et al (6). This method of analyzing MTR histograms is relatively simple: It is based on a comparison of a few arbitrary characteristics of MTR histograms, such as the position and height of the MTR peak between patient groups, and it enables the detection of changes between groups of patients.

Dehmeshki and co-workers (7,8) recently developed an alternative way of analyzing MTR histograms that has two major advantages compared with the existing method: First, it takes into account the entire shape of histograms rather than just a few arbitrary markers. Second, instead of performing t tests to compare groups of patients, one assigns patients to groups on the basis of their individual MTR histograms by using multivariate discriminant analysis (MDA). By using this method, patients with MS can be correctly assigned to clinical subgroups.

The aim of our preliminary study was to explore the diagnostic potential of MDA for assigning patients with NPSLE to different subgroups of patients with SLE on the basis of MTR histograms.


   MATERIALS AND METHODS
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
RESULTS
 DISCUSSION
 APPENDIX
 REFERENCES
 
Subjects
The following subjects were selected and underwent magnetization transfer imaging: 10 healthy female volunteers (mean age ± SD, 33 years ± 11), eight female patients and one male patient (mean age, 39 years ± 9) with active NPSLE, 10 female patients (mean age, 33 years ± 11) with past NPSLE, 10 female patients (mean age, 41 years ± 6) with MS, and 10 female patients with non-NPSLE. The characteristics, radiologic features, and neuropsychiatric manifestations in the patients with active and past NPSLE are listed in Table 1. All patients with SLE fulfilled the 1982 revised criteria for SLE of the American College of Rheumatology (9). The diagnosis of NPSLE was made on the basis of clinical findings and after the exclusion of other causes of the neuropsychiatric symptoms, as required by the American College of Rheumatology (10). To avoid the influence of thromboembolic processes on our study results, we excluded patients with radiologic evidence of infarctions other than incidental small (<5-mm) infarctions.


View this table:
[in this window]
[in a new window]

  		TABLE 1. Clinical Characteristics of Patients with Active or Past NPSLE and the Neuropsychiatric Manifestations at Onset

 
The patients with active NPSLE were selected from a group of consecutive patients participating in a treatment trial conducted at the Leiden University Medical Center, the Netherlands. For four of these patients, the neuropsychiatric symptoms at the time of imaging were the first manifestations of NPSLE. The patients with past NPSLE, non-NPSLE, or MS were selected on the basis of sex and age from our patient data files to match, in age and sex, the patients with active NPSLE. The patients with MS had clinically definite and laboratory-supported MS on the basis of the criteria used by Poser et al (11). Seven patients had relapsing-remitting MS, and three had the secondary progressive type of MS. At the time of imaging, none of the patients with MS had active disease. Approval from the local scientific committee of the departments of neurology, rheumatology, and radiology was obtained prior to commencement of the study, and all subjects gave informed consent before undergoing imaging.

Image Acquisition
Magnetic resonance (MR) imaging was carried out with a 1.5-T unit (Gyroscan; Philips Medical Systems, Best, the Netherlands). Conventional T1-weighted spin-echo MR images and intermediate- and T2-weighted turbo spin-echo MR images were obtained in all subjects and interpreted by an experienced neuroradiologist (M.A.V.B.). All sequences consisted of 22 6-mm-thick transverse sections, a 220-mm field of view, a 0.6-mm intersection gap, and a 256 x 205 matrix. T1-weighted MR imaging was performed by using a spin-echo sequence with 600/20 (repetition time msec/echo time msec) to result in an imaging time of 3 minutes 8 seconds. Intermediate-weighted imaging was performed by using a dual fast spin-echo imaging sequence with 2,500/23 or 120 and an echo train length of 10 to result in an imaging time of 2 minutes 35 seconds. Fluid-attenuated inversion-recovery fast spin-echo MR imaging was performed with 8,000/120 and an echo train length of 16 to produce the T2-weighted images in 3 minutes 28 seconds.

In addition to conventional MR imaging, all individuals underwent magnetization transfer imaging, which was performed by using a three-dimensional gradient-echo pulse sequence with 106/6 and a flip angle of 12°. These imaging parameters were chosen to minimize T1 and T2 weighting and thus resulted in contrast enhancement at intermediate-weighted imaging in the absence of magnetization transfer saturation pulses (12). A matrix of 256 x 128 pixels was used to obtain 28 sections with a thickness of 5 mm. The field of view was 220 mm. Two consecutive sets of transverse images were acquired: The first acquisition was performed in combination with a radio-frequency saturation pulse, and the second acquisition was performed without a radio-frequency pulse. The saturation pulse consisted of a sinc-shaped radio-frequency pulse 1,100 Hz down field of the water resonance with three periods of 15 msec each and an effective flip angle of 620°. The total imaging time for magnetization transfer imaging was 11 minutes 27 seconds.

Image Postprocessing
For postprocessing of magnetization transfer images, semiautomated software (3DVIEWNIX; Department of Radiology, Hospital of the University of Pennsylvania, Philadelphia) (13) was used at Leiden University Medical Center. A detailed description of our MTR analysis procedure has been given previously (14). In short, the following steps were performed: semiautomated segmentation of the intracranial volume from the three-dimensional magnetization transfer imaging study, automated calculation of MTRs for every intracranial voxel, and automated display of the three-dimensional data set of the MTRs as MTR histograms. MTR was defined as follows: MTR = [(M0 - Ms)/M0] x 100 pu, where Ms represents the signal intensity of voxels with saturation; and M0, the signal intensity of voxels without saturation. We used percent units (pu) as the unit of measure instead of percentage to avoid ambiguity (15). To reduce the partial volume effects from brain surface voxels that included both brain tissue and cerebrospinal fluid, we excluded those voxels with MTRs of less than 20 pu. Subsequently, we corrected the MTR histograms for intracranial volume by dividing the individual bin values by the total number of intracranial voxels. Thus, the total area under the histogram curve was fixed at unity, and this allowed a comparison of histograms originating from individuals with different intracranial volumes. From each intracranially normalized histogram, the MTR peak height, MTR corresponding to the peak (ie, MTR peak position), and mean MTR were calculated.

MTR Histogram Classification
A detailed description of a general approach to MDA is provided in the Appendix (Eqq [A1]–[A7]) and has been reported previously (7,8). In this section, we summarize the MDA procedures used to achieve the fully automated system for classification of the MTR histograms that was implemented at the Institute of Neurology, University College London, England.

The aim of MDA is to maximize the ratio of the between-group variance to the within-group variance. This is achieved by finding a set of coefficients that, when multiplied by each value in the histogram, give a score that optimally discriminates between the subgroups under consideration. These coefficients are different for each binary comparison. Thus, the MDA score contains information from the whole histogram in a way that has been optimized to place the most emphasis on those parts that contribute most to the separation of the groups. This approach is an improvement of that involving the use of single parameters, such as peak height or peak position, to characterize a histogram. The MDA score is then used to decide which group an individual histogram belongs to. Thus, the first group in a binary comparison (eg, control subjects) has a positive mean score, and the second group (eg, patients) has a negative mean score.

For a particular subject who is being assigned to a group, if the MDA score is positive, then the subject is assigned to the second group; otherwise, subjects are assigned to the first group. Depending on the amount of overlap between subgroups, there may be some errors in this classification procedure. Examples of successes and failures in our classification are described in the Results section. Because the sizes of the training sets were small (only 10–12 sets in each group), we used the following assumption strategies in our implementation:

Leave-one-out method for testing.—Ideally, a classifier will be evaluated by means of a test set of individuals of unknown clinical groups that is distinct from the training set. However, this evaluation of the classifier requires a large total data set. We used a leave-one-out method to create the test data. In this procedure, each individual histogram, which represents the test individual, is in turn removed from the total number of subjects in the two groups (ie, N individuals) being analyzed. This left-out histogram is then classified by using the MDA trained by the histograms from the other N-1 individuals. The total number of misclassifications is then an almost unbiased estimate of the expected true error rate (16,17).

Equal covariance matrixes.—It is better to assume an equal covariance matrix for all groups. This is because a small number of samples cannot guarantee a good estimation of the class covariance (18).

Binary comparison.—We evaluated the system by attempting to distinguish between pairs of clinical groups. Although this binary approach might seem artificial, it is helpful in clinical practice, because frequently, in patients with neurologic and psychiatric symptoms who have or might have SLE, two possible conditions have to be discriminated. (For more details, see Discussion section.) The actual classification procedure that we used, with the restrictions just described taken into account, is as follows:

1. An individual histogram—that is, the test sample—was selected from the N individual samples in the two subgroups.

2. The linear discriminant coefficients (ie, weights) that best discriminated between two clinical groups, as described in the Appendix, were derived from the training set histogram samples—that is, the N-1 samples.

3. The discriminant score for the test histogram was calculated by multiplying the histogram values by the discriminant coefficients obtained in step 2.

4. A nearest-mean-class classifier technique was used to assign the test histogram to one of the two subgroups by measuring the distance between the test histogram’s particular score and the mean scores of the two groups.

Steps 1–4 were repeated for the MTR histogram of each individual sample.

Comparison of MDA Scores versus Conventional Histogram Parameters
To facilitate a convincing comparison of our MDA score method with the method involving conventional histogram parameters (ie, peak height, peak position, and mean value), Student t tests were carried out between some subgroups. The results of previous work (4) have shown significant differences between patients with active SLE and individuals in other subgroups, particularly control subjects. We investigated whether MDA scores could facilitate a better separation between the groups (ie, lower P values) than conventional histogram parameters.


   RESULTS
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
RESULTS
DISCUSSION
 APPENDIX
 REFERENCES
 
The mean MTR histograms for the four clinical groups—namely, active NPSLE, past NPSLE, non-NPSLE, and MS—and for the control subjects are shown in Figure 1. The distribution of discriminant scores for each of the 10 binary classifications is shown in Figure 2. Scatterplots of the conventional parameters (Fig 3) show increased overlap compared with the distribution of MDA scores. The superior performance of the MDA scores is confirmed by the P values presented in Table 2, where P values for the ability to distinguish the groups, as measured by using t tests, are shown for conventional histogram parameters and MDA scores. In two of the four comparisons shown, peak position was the best of the three conventional parameters for separating the groups, and in the other two cases, mean MTR was the best. In contrast, MDA was always better than any of these parameters, with P values ranging from .05 to less than 1 x 10-6 of those attained with the best conventional parameter.



View larger version (47K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 1. Mean histograms for the active NPSLE (AC), control (CO), MS, past NPSLE (PN), and non-NPSLE (NON) disease groups. These histograms are normalized such that the total area under the histogram curve is fixed at unity.

 


View larger version (20K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 2. Scatterplots for the 10 binary classifications. The vertical axis shows the discriminant MDA score. Some groups, such as the active NPSLE () and control ({square}) groups, are well separated with no classification errors, whereas others, such as the past NPSLE ({bullet}) and MS ({triangleup}) groups, have considerable overlap and classification errors. {circ} = non-NPSLE group.

 


View larger version (12K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 3a. Scatterplots illustrate comparisons between the active NPSLE group () and the other groups, performed by using the following conventional parameters: (a) mean MTR, (b) MTR peak height, and (c) MTR peak location. In c, several samples have the same (integer) peak location value and thus coincide. {square} = control group, {triangleup} = MS group, {circ} = non-NPSLE group, {bullet} = past NPSLE group.

 


View larger version (14K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 3b. Scatterplots illustrate comparisons between the active NPSLE group () and the other groups, performed by using the following conventional parameters: (a) mean MTR, (b) MTR peak height, and (c) MTR peak location. In c, several samples have the same (integer) peak location value and thus coincide. {square} = control group, {triangleup} = MS group, {circ} = non-NPSLE group, {bullet} = past NPSLE group.

 


View larger version (12K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 3c. Scatterplots illustrate comparisons between the active NPSLE group () and the other groups, performed by using the following conventional parameters: (a) mean MTR, (b) MTR peak height, and (c) MTR peak location. In c, several samples have the same (integer) peak location value and thus coincide. {square} = control group, {triangleup} = MS group, {circ} = non-NPSLE group, {bullet} = past NPSLE group.

 

View this table:
[in this window]
[in a new window]

  		TABLE 2. MDA Score and Conventional Histogram Parameter P Values in Distinguishing Active NPSLE Group from Other Disease Groups

 
Examples of how the discriminant P values were used for the binary classification technique are shown in Figure 4. The number of correctly classified histogram samples and the total number of samples involved in each binary comparison are given in Table 3. The data show that it was relatively difficult to distinguish the past NPSLE group from the MS group: There were 12 successes in 20 tests (Table 3). On the other hand, the active NPSLE group was well distinguished from all other groups: There were four failures in 76 tests—two in the active NPSLE versus past NPSLE tests and two in the active NPSLE versus MS tests (Table 3).



View larger version (13K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 4. Graph illustrates examples of binary classification. One sample is treated as an unknown subject for training with the remaining samples. On the left, in the binary differentiation between the active NPSLE group () and the control group ({square}), the unknown subject ({circ}), which in fact represents a patient with active NPSLE, has an MDA score that is far from zero (score, -7.54). Thus, we can be sure of the classification, and, indeed, the subject was correctly assigned to the active NPSLE group. In the middle, in the binary differentiation between the active NPSLE group and the past NPSLE group ({bullet}), the unknown subject, which in fact also represents a patient with active NPSLE (score, -1.41), has an MDA score that is nearer to the mean score of the active NPSLE group. This low score and the position in the region of overlap tell us that confidence in the classification was low. Nonetheless, the subject was correctly assigned. On the right, in the binary differentiation between the active NPSLE group and the past NPSLE group, the unknown subject, which in fact represents a patient with past NPSLE (score, -0.42), has an MDA score that is nearer to the mean score of the active NPSLE group. Confidence in the classification was low, however, and the subject was incorrectly assigned.

 

View this table:
[in this window]
[in a new window]

  		TABLE 3. Rates of Correct Classification at Binary Classification

 

   DISCUSSION
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
RESULTS
 DISCUSSION
APPENDIX
 REFERENCES
 
In a recent study by Bosma and co-workers (5), differences between a past NPSLE group and a non-NPSLE group were found when MTR histogram analysis was used. The differences included decreased MTR histogram peak height and increased MTR SDs, both of which were reflective of a global loss of cerebral homogeneity in the past NPSLE patient group. In a subsequent study by the same investigators (4), a difference between an active NPSLE group and a past NPSLE group was found (4). This difference was a marked shift of the MTR histogram peak to a higher MTR, which indicated a global increase in the magnetization transfer effect in the brain. This increase probably was based on a diffuse process in the active phase of the disease that differs from the residual changes that persist after the regression of symptoms. The MTR changes seen in the active stage of NPSLE are similar to those seen in the acute stage of wallerian degeneration and experimental allergic encephalitis and might reflect transient inflammatory changes (12,19). The MTR changes seen in patients with past NPSLE are similar to those seen in patients with demyelinating diseases such as MS and might reflect residual structural changes such as gliosis and demyelination (4).

The method of analyzing volumetric MTR histograms that has been used thus far has been limited (12,20–28). Arbitrarily chosen descriptive measures were used to describe the MTR histogram and to measure changes. These measures included MTR histogram peak height; peak MTR; and 25th, 50th, and 75th percentile values. This method enabled the detection of cerebral differences between patient groups by means of simple t tests.

A new approach to analyzing MTR histograms was recently developed by Dehmeshki and co-workers (7,8). This approach has two major advantages over the existing method of analyzing MTR histograms: (a) It is a more general approach to the global characterization of the histogram that potentially involves the use of the entire histogram, and (b) it provides a way of assigning an individual patient to a disease group on the basis of his or her volumetric cerebral MTR histogram. In short, this MDA approach consists of finding a set of coefficients that, when multiplied by each value in the histogram, give a score that optimally discriminates between the subgroups under consideration. These coefficients are different for each binary comparison. This score is then used to decide which group an individual histogram belongs to. Depending on the amount of overlap between the subgroups, there may be some errors in this classification procedure.

In the preliminary study described herein, only small groups of 10–12 patients with a certain diagnosis could be used for training the system, and, consequently, the diagnostic power of MDA was not optimal. For this reason, we chose to make binary comparisons between pairs of clinical groups. Although this approach might seem artificial, in the clinical setting involving patients with NPSLE, decisions are often binary. From a clinical perspective, the following three binary comparisons are relevant because the diseases in each comparison require different treatment:

First, a comparison of active NPSLE versus past NPSLE groups is important because in patients with SLE and a history of neurologic and psychiatric symptoms, a new episode of neuropsychiatric symptoms could be due to an active intrinsic SLE-related brain process—that is, a recurrent active phase of NPSLE—or to extrinsic processes such as the side effects of drug use or the consequences of antiphospholipid antibodies (an extrinsic process in patients with past NPSLE). Second, distinguishing active NPSLE from other types of SLE is important because a de novo acute episode of neuropsychiatric symptoms could be due to active NPSLE or to extrinsic processes in patients with SLE. Third, it is important to compare patients who have active and/or past NPSLE with those who have MS because in patients who present with neuropsychiatric symptoms, the differential diagnosis includes SLE and MS, and these diseases are notoriously difficult to differentiate.

In this preliminary study, MDA proved to be effective in assigning the majority of individual patients to disease categories in the given binary comparisons. In addition, MDA was shown to be considerably more effective for categorizing patient groups on the basis of MTR histograms than the conventional method of analyzing MTR histograms. The data presented in this study are promising and suggest that MDA of MTR histograms has diagnostic potential that might be of help in clinical practice.

Larger studies are needed to further investigate the diagnostic potential of MDA of MTR histograms. One limitation of the present study was the relatively small size of the patient subgroups. Larger numbers of patients will permit more thorough training of the system, which might increase the diagnostic power of the method. Another limitation of this study was that the patient groups included subjects with unambiguous diagnoses. In clinical practice and for reasons discussed earlier, diagnoses often are ambiguous; thus, the classification of such cases might be more difficult. Larger studies with a wider spectrum of cases are needed to further explore the diagnostic potential of MDA of MTR histograms.

In summary, the current study represents an investigation of an analysis method that can be used to assess the condition of the brain on the basis of a global characterization of MTR histograms and that proved successful in classifying different categories of patients with SLE. To our knowledge, this is the first description of a method that enables the use of magnetization transfer data for diagnostic purposes in individual patients. The data in this preliminary study are promising; however, with a larger training set and the consequent use of nonlinear discriminant analysis, we expect to see further improved performance.


   APPENDIX
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
RESULTS
 DISCUSSION
 APPENDIX
REFERENCES
 
The following is a brief summary of MDA, but greater detail regarding this analysis method is described elsewhere (7,8,18). The aim of MDA is to maximize the ratio of between-group variance to within-group variance:

where Wb is the between-class covariance matrix; Ww, the within-class covariance; and {phi}, the transformation we are searching for to form the optimal discriminant space.

The mean MTR histogram for each patient group is defined as follows:

for the group mean histogram ( ) of ni patients (p) in group i.

An N x N covariance matrix is then calculated for each group as follows:

for the covariance matrix of group i (Wi) with elements k,l = 0,...,N-1, where N is the histogram bin size. In addition, the following calculations can be performed:

where n is the total number of samples.

In looking for {phi}, we can define the following:

It follows from these equations that

Taking the determinant of a covariance matrix is equivalent to finding the product of the eigenvalues, which corresponds to the product of the variance. To maximize the ratio in Equation (A1), we look for a transform {phi} that maximizes the between-class variance with respect to the within-class variance. The solution to Equation (A1) can be shown to correspond to the generalized eigenvectors of the following equation:

where the {phi}i vectors are the columns of the matrix {phi}.


   FOOTNOTES
 
Abbreviations: MDA = multivariate discriminant analysis, MS = multiple sclerosis, MTR = magnetization transfer ratio, NPSLE = neuropsychiatric SLE, SLE = systemic lupus erythematosus

Author contributions: Guarantor of integrity of entire study, J.D.; study concepts and design, J.D., M.A.V.B., P.S.T.; literature research, J.D., G.P.T.B.; clinical studies, G.P.T.B., T.W.J.H.; data acquisition, G.P.T.B., T.W.J.H.; data analysis/interpretation, J.D., M.A.V.B., G.P.T.B.; statistical analysis, J.D.; manuscript preparation, J.D., G.P.T.B., M.A.V.B.; manuscript definition of intellectual content and editing, J.D.; manuscript revision/review, P.S.T., M.A.V.B., T.W.J.H.; manuscript final version approval, J.D., M.A.V.B., T.W.J.H., P.S.T.


   REFERENCES
TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
RESULTS
 DISCUSSION
 APPENDIX
 REFERENCES
 

  1. Johnson RT, Richardson EP. The neurological manifestations of systemic lupus erythematosus. Medicine (Baltimore) 1968; 47:337-369.
  2. Devinsky O, Petito CK, Alonso DR. Clinical and neuropathological findings in systemic lupus erythematosus: the role of vasculitis, heart emboli, and thrombotic thrombocytopenic purpura. Ann Neurol 1988; 23:380-384.
  3. West SG. Neuropsychiatric lupus. Rheum Dis Clin North Am 1994; 20:129-158.
  4. Bosma GPT, van Buchem MA, Rood MJ, de Jong BA, Bollen ELEM, Huizinga TWJ. Detection of cerebral involvement in patients with active neuropsychiatric systemic lupus erythematosus using volumetric magnetization transfer imaging. Arthritis Rheum 2000; 43:2428-2436.
  5. Bosma GPT, Rood MJ, Zwinderman AH, Huizinga TWJ, van Buchem MA. Evidence of CNS damage in patients with neuropsychiatric systemic lupus erythematosus demonstrated by magnetization transfer imaging. Arthritis Rheum 2000; 43:48-54.
  6. Van Buchem MA, Udupa JK, McGowan JC, et al. Global volumetric estimation of disease burden in multiple sclerosis based on magnetization transfer imaging. AJNR Am J Neuroradiol 1997; 18:1287-1290.
  7. Dehmeshki J, Silver N, Ruto AC, Arridge S, Miller DH, Tofts PS. Detection and classification of multiple sclerosis disease using magnetisation transfer ratio images In: Proceedings of the 1st International Workshop on Image and Signal Processing and Analysis IWISPA 2000, 22nd International Conference on Information Technology Interfaces. Pula, Croatia: International Conference on Information Technology Interfaces, 2000; 127-132.
  8. Dehmeshki J, Ruto AC, Arridge S, Silver NC, Miller DH, Tofts PS. Analysis of MTR histograms in multiple sclerosis using principal components and multiple discriminant analysis. Magn Reson Med 2001; 46:600-609.
  9. Tan EM, Cohen AS, Fries JF, et al. The 1982 revised criteria for the classification of systemic lupus erythematosus. Arthritis Rheum 1982; 25:1271-1277.
  10. The American College of Rheumatology nomenclature and case definitions for neuropsychiatric lupus syndromes. Arthritis Rheum 1999; 42:599-608.
  11. Poser CM, Paty DW, Scheinberg L, et al. New diagnostic criteria for multiple sclerosis: guidelines for research protocols. Ann Neurol 1983; 13:227-231.
  12. Dousset V, Grossman RI, Ramer KN, et al. Experimental allergic encephalomyelitis and multiple sclerosis: lesion characterization with magnetization transfer imaging. Radiology 1992; 182:483-491[Erratum: Radiology 1992; 183:878.].
  13. Udupa J, Odhner D, Samarasekera S, et al. 3DVIEWNIX: an open, transportable, multidimensional multimodality, multiparametric imaging software system. SPIE Proc 1994; 2164:58-73.
  14. Van Buchem MA, Udupa JK, McGowan JC, et al. Global volumetric estimation of disease burden in multiple sclerosis based on magnetization transfer imaging. AJNR Am J Neuroradiol 1997; 18:1287-1290.
  15. Barker GJ, Tofts PS, Gass A. An interleaved sequence for accurate and reproducible clinical measurement of magnetization transfer ratio. Magn Reson Imaging 1996; 4:403-411.
  16. Fukunaga K, Hayes R. Estimation of classifier performance. IEEE Trans Pattern Analysis and Machine Intell 1989; 11:1087-1097.
  17. Hills M. Allocation rules and their error rates (with discussion). J Roy Stat Soc 1966; 28:1-31.
  18. Webb A. Statistical pattern recognition New York, NY: Oxford University Press, 1999; 6.
  19. Lexa FJ, Grossman RI, Rosenquist AC. Dyke award paper: MR of wallerian degeneration in the feline visual system—characterization by magnetization transfer rate with histopathologic correlation. AJNR Am J Neuroradiol 1994; 15:201-212.
  20. Hiehle JF, Jr, Grossman RI, Ramer KN, et al. Magnetization transfer effects in MR-detected multiple sclerosis lesions: comparison with gadolinium-enhanced spin-echo images and nonenhanced T1-weighted images. AJNR Am J Neuroradiol 1995; 16:69-77.
  21. Van Buchem MA, Udupa JK, McGowan JC, et al. Global volumetric estimation of disease burden in multiple sclerosis based on magnetization transfer imaging. AJNR Am J Neuroradiol 1997; 18:1287-1290.
  22. Rovaris M, Filippi M, Falautano M, et al. Relation between MR abnormalities and patterns of cognitive impairment in multiple sclerosis. Neurology 1998; 50:1601-1608.
  23. Erickson BJ, Noseworthy JH. Value of magnetic resonance imaging in assessing efficacy in clinical trials of multiple sclerosis therapies. Mayo Clin Proc 1997; 72:1080-1089.
  24. Udupa JK, Wei L, Samarasekera S, et al. Multiple sclerosis lesion quantification using fuzzy-connectedness principles. IEEE Trans Med Imaging 1997; 16:598-609.
  25. Van Buchem MA, McGowan JC, Kolson DL, et al. Quantitative volumetric magnetization transfer analysis in multiple sclerosis: estimation of macroscopic and microscopic disease burden. Magn Reson Med 1996; 36:632-636.
  26. Petrella JR, Grossman RI, McGowan JC, et al. Multiple sclerosis lesions: relationship between MR enhancement pattern and magnetization transfer effect. AJNR Am J Neuroradiol 1996; 17:1041-1049.
  27. Mehta RC, Pike GB, Enzmann DR. Measure of magnetization transfer in multiple sclerosis demyelinating plaques, white matter ischemic lesions, and edema. AJNR Am J Neuroradiol 1996; 17:1051-1055.
  28. Loevner LA, Grossman RI, McGowan JC, et al. Characterization of multiple sclerosis plaques with T1-weighted MR and quantitative magnetization transfer. AJNR Am J Neuroradiol 1995; 16:1473-1479.



This article has been cited by other articles:


Home page
 J. Neurol. Neurosurg. PsychiatryHome page
B. Benedetti, M. Rovaris, E. Judica, G. Donadoni, G. Ciboddo, and M. Filippi
Assessing "occult" cervical cord damage in patients with neuropsychiatric systemic lupus erythematosus using diffusion tensor MRI
J. Neurol. Neurosurg. Psychiatry, August 1, 2007; 78(8): 893 - 895.
[Abstract] [Full Text] [PDF]

Home page
 LupusHome page
J W Graham and W Jan
MRI and the brain in systemic lupus erythematosus
Lupus, December 1, 2003; 12(12): 891 - 896.
[Abstract] [PDF]

Home page
 LupusHome page
P L Peterson, F A Howe, C A Clark, and J S Axford
Quantitative magnetic resonance imaging in neuropsychiatric systemic lupus erythematosus
Lupus, December 1, 2003; 12(12): 897 - 902.
[Abstract] [PDF]

This Article
Right arrow Abstract Freely available
Right arrow Figures Only
Right arrow Full Text (PDF)
Right arrow All Versions of this Article:
2223010413v1
222/3/722    most recent
Right arrow Submit a response
Right arrow Alert me when this article is cited
Right arrow Alert me when eLetters are posted
Right arrow Alert me if a correction is posted
Right arrow Citation Map
Services
Right arrow Email this article to a friend
Right arrow Similar articles in this journal
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Dehmeshki, J.
Right arrow Articles by Tofts, P. S.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Dehmeshki, J.
Right arrow Articles by Tofts, P. S.

HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
RADIOLOGY RADIOGRAPHICS RSNA JOURNALS ONLINE