Comparison of different MRI brain atrophy rate measures with clinical disease progression in AD

Recruitment and clinical characterization.

Patients and control subjects were recruited from the Mayo Alzheimer’s Disease Patient Registry (ADPR) and Alzheimer’s Disease Research Center (ADRC).14 These longitudinal studies of aging and dementia provide an ongoing mechanism for identification, enrollment, and longitudinal clinical characterization of both community and referral patients. Both the ADPR and the ADRC include MRI studies. These studies were performed with Mayo Institutional Review Board approval and informed consent of the subject or an appropriate proxy. At both baseline and follow-up assessments, the subjects must have been classified into one of three clinical groups: cognitively normal, MCI, or probable AD. Potential subjects who had cognitive symptoms that were clinically felt to be unrelated to AD were excluded. For example, subjects who experienced a stroke or who developed depression before or during the follow-up period were excluded. Study participants were assigned to diagnostic group categories during ADPR/ADRC Consensus Committee meetings consisting of a geriatrician, neurologist, neuropsychologist, psychometrist, and nurses who had seen the patient. The same Consensus Committee members assigned diagnostic group categories to patients at both the first and the second clinical follow-up point.

Categorization into diagnostic groups represented a clinical judgment that was based on the combined results of medical history, clinical exam, and psychometrics. Criteria for the diagnosis of normal cognition were as follows: 1) no active neurologic or psychiatric disorders, 2) any ongoing medical problems or their treatments did not interfere with cognitive function, 3) a normal neurologic exam, 4) no psychoactive medications, 5) independently functioning community dwellers. Criteria for the diagnosis of MCI were as follows: 1) memory complaint documented by the patient and collateral source, 2) normal general cognition, 3) normal activities of daily living, 4) no dementia (Diagnostic and Statistical Manual of Mental Disorders [3rd rev. ed.; DSM-IIIR]), 5) memory impairment for age and education.15 The diagnosis of MCI was a clinical categorization of individuals who displayed a memory impairment with relative preservation of other cognitive functions, that is, amnestic MCI.15–18⇓⇓⇓ The diagnosis of probable AD was made according to the DSM-IIIR criteria for dementia and National Institute of Neurologic and Communicative Disorders and Stroke/Alzheimer’s Disease and Related Disorders Association criteria for AD.19,20⇓

Clinical transition analysis.

In addition to the baseline diagnostic group assignment, each normal and MCI subject was further classified as either stable or converter. Subjects were designated as stable if their clinical classification assigned by the ADPR/ADRC Consensus Committee did not change; that is, normal subjects remained normal, and MCI subjects remained classified as MCI. Subjects were designated as converters if their clinical classification changed; that is, normal subjects converted to MCI or AD, or MCI subjects converted to AD. The AD group was dichotomized into slow progressor or fast progressor groups. A slow progressor was defined as an individual who did not decline on the Clinical Dementia Rating Scale (CDR) summary score.21 A fast AD progressor was defined as an individual who did decline on the CDR summary score. For the clinical transition analyses, therefore, each subject was assigned to one of six possible groups: normal stable, normal converter, MCI stable, MCI converter, AD slow progressor, or AD fast progressor.

All subjects in the ADPR/ADRC are followed with approximately annual clinical assessments. The interval between first and last MRI scans varied by baseline clinical group. The follow-up interval range for subjects who were classified as normal at baseline was 2.2 to 5.2 years; for subjects classified as MCI at baseline, 1.2 to 4.8 years; and for subjects classified as AD at baseline, 0.9 to 2.4 years.

MRI methods.

All MR studies were performed within 4 months of the matching clinical/psychometric assessment. The maximum interval between MRI and linked clinical assessment was 2 months in 131 subjects, 3 months in 22 subjects, and 4 months in 7 subjects. Results of the MRI studies were used clinically only to identify treatable causes of dementia. The numeric data on brain atrophy rates were produced for research purposes only, were unknown to the Consensus Committee throughout the study, and were not employed in the clinical classification process.

All patients were imaged at 1.5 T (Signa; General Electric Medical Systems, Milwaukee, WI) using a standardized imaging protocol. All volume measurements were derived from a T1-weighted three-dimensional volumetric spoiled gradient-recalled echo sequence with 124 continuous partitions, 1.6-mm slice thickness, a 22 × 16.5–cm field of view, 192 views, and 25° flip angle.

All image processing was performed by the same research associate who was blinded to all clinical information (e.g., age, gender, and clinical status). The date of each MR scan was masked in the image header file so that the imaging measures were done without knowledge of the chronologic ordering of the scans in each scan pair. The rates of atrophy for each of the four MR measures were expressed as percentage change in volume divided by the interval between the scans in years.

Hippocampal and ERC measurements were performed by manual tracing after several image-preprocessing steps had been performed.22 The three-dimensional image data set of the second scan was spatially registered with that of the first scan. The image data of both scans were then interpolated in-plane to the equivalent of a 512 × 512 matrix and magnified by 2. The images of the whole brain were then subvolumed to include the temporal lobes. An intensity inhomogeneity correction algorithm developed in-house was then applied to both MR scans. Tracing of the paired scans for each subject was performed side by side. After hippocampal and ERC boundaries had been delineated on each anatomic slice, the number of voxels was calculated automatically with a summing region-of-interest function. These were multiplied by voxel volume to give a numeric value in cubic millimeters.

The borders of the right and left hippocampi were manually traced for each slice sequentially from posterior to anterior. In-plane hippocampal anatomic boundaries were defined to include the CA1 through CA4 sectors of the hippocampus proper, the dentate gyrus, and subiculum.23 The posterior boundary of the hippocampus was determined by the oblique coronal anatomic section on which the crura of the fornices were identified in full profile.

We employed slight modifications of criteria published by Insausti et al. and Killiany et al. for determining boundaries of the ERC on MRI.24–26⇓⇓ Its inferolateral boundary was defined by the medial edge of the collateral sulcus posteriorly and by the rhinal sulcus anteriorly. In some cases, the sulcus was bi- or multilobed. In these sections, the more medially located sulcus was used as the border. The medial boundary was CSF in the crural cistern. The ERC was demarcated from the overlying subiculum by a medial extension of the horizontal line defined by the gray–white subicular interface. The posterior boundary of the ERC was defined as the first imaging section containing the intralimbic gyrus when progressing from posterior to anterior. In-plane boundaries were measured from this slice forward on six consecutive sections bilaterally.

Rates of change of the whole brain and ventricle were made using a home-built software algorithm, described in detail elsewhere.27 Inputs for this algorithm are two or more three-dimensional volumetric MRI scans obtained at different times. The algorithm begins with a series of data-preprocessing steps including extracting the brain from the overlying skull and scalp, creating binary masks of brain and ventricle, correcting B₁ field nonuniformity, spatial alignment of the two image volumes, scaling correction, and intensity normalization. The preprocessed brain volumes are then entered into a module that computes change in brain and ventricular volume from baseline to follow-up MRI exam. This volume change computation method was modified from the SIENA28 software package (University of Colorado, Boulder). Volume change is determined at the brain–CSF interface over the entire three-dimensional surface of the registered brain volumes. Computation of volume change is conceptually very similar to the boundary shift integral method of Freeborough and Fox29 except that volume change is derived from a measure of the intensity gradient perpendicular to the brain–CSF interface in the registered MR volumes rather than from the absolute intensity difference.

Psychometric assessment.

Psychometric assessments within 4 months of MRI scans were obtained for 89% of normal, 98% of MCI, and 100% of AD subjects. We selected a limited number of tests from a more extensive testing battery for MRI–cognitive correlation analyses to appropriately avoid floor and ceiling effects. The Mini-Mental State Examination (MMSE)30 and Dementia Rating Scale (DRS) were employed for all subjects—normal, MCI, and AD—to provide continuity across all subjects in the study. The MMSE, although somewhat insensitive to change at the upper end of the cognitive range, was selected because it is widely used in the field. The DRS, although less widely used, has a greater dynamic range than the MMSE. The CDR was originally designed to stage disease severity in dementia, and we performed the correlations with the Sum of Boxes score only in the MCI and AD groups. For memory assessment, we employed the Auditory Verbal Learning Test (AVLT)–Delayed Recall and the percentage retention on the Logical Memory II Subtest of the Wechsler Memory Scale–Revised (LM II [%]).31,32⇓ The AVLT–Delayed Recall is a relatively difficult test, and therefore AVLT–MRI correlations were performed only in the cognitively normal baseline cohort. The LM II (%) is less difficult than the AVLT and was used to provide a memory measure common to the normal and MCI groups. The change from baseline to follow-up performance was measured in each subject and annualized.

Statistical methods.

The use of different intervals between baseline and follow-up for individuals in each of the three baseline clinical groups was an intentional feature of the study design. Rates of clinical transition are an order of magnitude greater for individuals with MCI vs cognitively normal subjects: 12 to 15%/year of individuals with MCI will decline to AD, whereas only 1 to 2%/year of normal subjects will decline to MCI.16,33⇓ To observe a comparable number of conversion events, either the number of normal subjects enrolled would have to be an order of magnitude greater than the number of MCI subjects or the follow-up interval would have to be greater for normal vs MCI subjects. We chose the latter approach. We did ensure, however, that the follow-up interval was approximately the same between control vs stable members of the normal and MCI cohorts and between fast- vs slow-progressing AD subjects.34 The rate comparisons of interest were within members of the three baseline clinical groups.

The distributions of many of the MRI and psychometric variables were skewed. Therefore, for the sake of consistency, all summary statistics for quantitative variables are reported as median values and variability as interquartile range (IQR). All tests for significance were performed using nonparametric methods and were two sided. For both MRI and cognitive measures, tests for differences in rates of change between stable vs converter normal and MCI subjects and between fast vs slow AD progressor subjects were performed using rank sum tests. Correlations between change in MRI volumes and change in cognitive scores were tested for significance using Spearman correlations.

To test whether the strength of the association between rate of atrophy and clinical transition status (stable vs converter, fast vs slow for AD) differed among the four MR measurements, we used the basic approach proposed by Choi35 separately for each cohort (normal, MCI, AD). To illustrate using the normal cohort, the dependent variable was defined as 1 if stable and 0 if converter. Four independent variables (x₁, x₂, x₃, and x₄) were defined, one for each MR rate measurement, as the ranks of the volumes in the pooled (stable plus converter) cohort. Defining z₁ = x₁ − x₂, z₂ = x₁ + x₂ − 2x₃, and z₃ = x₁ + x₂ + x₃ − 3x₄, we tested that clinical transition status was unrelated to the three z variables using logistic regression. This is equivalent to testing that the association between clinical transition status and rate of atrophy is the same among the four MRI measurements.35 To compare the associations between only two atrophy measurement techniques, only z₁ is defined, and a rank sum test was used to compare z₁ values between stable and converter subgroups. This approach tests the null hypothesis that the distribution overlap between the two subgroups (stable vs converter, e.g.) is the same for each MRI rate measurement.

The ability of longitudinal change in volume to distinguish between stable vs converter (or fast vs slow AD) subgroups depends on the overlap between the two subgroups, and this in turn depends on both the differences in the medians and the variability in the data. It is desirable to have an index measuring the overlap between groups for each MRI rate measurement. For this purpose, we performed a logistic regression discriminating between the two subgroups for each MRI measurement. Overlap may be measured by the odds ratio (OR) from the logistic regression. The independent variable was defined as the percentile of the annual percentage change (APC) in the pooled (stable plus converter subgroups) data set. The resulting OR per decile measures the overlap between stable vs converter (or fast vs slow AD) subgroups.

Recognizing general interest in sample size estimates for therapeutic trials using imaging, we provided power calculations for 25 and 50% treatment effect sizes in the MCI group. Effect size represents a 25 or 50% reduction in the median transformed rates of MRI atrophy or change in cognitive test/rating scale performance. For a given MRI or cognitive measure, we calculated the median observed value among all MCI subjects as well as the median we would expect to see under a 50 or 25% improvement in the rate of decline. We next found a transformation that would make the observed data approximately normal and applied this transformation to the observed median and the improvement median. The difference between these two values became the effect size we sought to detect with 90% power using an unpaired two-sided t-test. To estimate the population standard deviation on the scale used for the t-test, we calculated the sample standard deviation of the transformed data. The transformations used were of the form y = (x + c)^p. Here, x is the observed rate of change, c is the smallest integer ≥0 such that x + c is positive for all x, and p = 1/2 when the data were right skewed or p = 2 when the data were left skewed. As the MMSE measurements were approximately normal, observed means were used rather than medians to calculate effect size, and no transformation was performed.

An understanding of the precision of the four different MR measurements is helpful in interpreting the results. Most of the published data on this topic address the precision of repeated measures of the volume of a particular brain structure (e.g., hippocampal volume), which is not the issue here. We are interested in the test–retest reproducibility of measurements of change (i.e., reproducibility of Δ measurements). To provide these data, three subjects were randomly selected from each of the six clinical groups in this study. These 18 subjects had all MR measurements repeated 1 year later by same research associate, who was blinded to results of the first measurements. The difference in APC as measured initially minus the APC as measured 1 year later was calculated. We refer to these differences as Δ values. Precision was measured two ways. First, we computed the difference (Δ) between the two APC values for each subject and the corresponding IQR for each of the four MRI measures. To test for differences among MRI measures, we performed the Friedman two-way analysis of variance on the ranks of the absolute values of the differences. Pairwise comparisons were performed using signed rank tests on the absolute differences. With the second approach, we calculated intraclass correlations (ICC) based on the ranks of the 18 original + 18 retest = 36 APC. This nonparametric ICC represents the proportion of total variability in the ordering of the measurements that is due to person-to-person variability: (1 − ICC) is the proportion of total variability due to test–retest variability.

Each subject’s baseline MRI volumes served an as internal control for intersubject variation in head size. Therefore, there was no need to perform brain volume normalization when comparing rates among different subjects.34