Medical imaging for patient care invokes considerations at the levels of the individual patient and the individual imaging site. However, medical imaging in clinical research studies is often performed at multiple sites. This is true for large observational studies and is particularly true for therapeutic trials, which are nearly always conducted at multiple sites. The reasons for the this are fairly obvious. If one wishes to capture the relevant variation present across an entire population for a natural history study, then subjects must be recruited from multiple sites. The same considerations hold for therapeutic trials in which therapeutic efficacy across a representative range of the population must be demonstrated. In addition, recruiting the number of subjects needed to power a study is often impossible at a single site.
Like other imaging modalities, MRI captures morphometric or functional data that provide useful information about pathological processes of relevance. That is, imaging measures serve as in vivo surrogates of relevant pathologies. Because of its flexible nature, MRI can provide information about a variety of anatomical and physiological brain processes. These processes include brain morphology, changes in relaxation properties, perfusion, diffusion, and metabolite concentration.
Variability in imaging data collected across different subjects or across individual subjects over time can be considered in three categories. First is the data variability due to the effect of the pathology one seeks to measure. For example, variability in brain volume may be due to presence and severity of AD or due to TBI. Second is data variability due to biology that is irrelevant to the pathology of interest. For example, brain volume may vary with hydration or nutritional status. Third is data variability due to technical or engineering-related factors. In any study, whether single site or multisite, the objective is to maximize the impact of biologically relevant data variability and minimize the impact of other sources of variability. The reason for this objective is self-evident; the more variability in a data set that is directly due to the pathology of interest, the more useful the imaging will be in probing the biologically relevant relationships. Conversely, irrelevant biological variability and engineering-related variability will only obscure the relationships between imaging and the pathology of interest. In designing multisite trials, unwanted data variability due to the irrelevant biology can be minimized to some extent by rigorous inclusion and exclusion criteria. But the source of undesirable data variability that is under greatest control is that due to technical or engineering-related factors.
A variety of specific items should be considered in the design of multisite trials. However, an overarching principle is that unwanted data variability due to technical factors can be minimized by standardization. And the principle of standardization applies in two dimensions: (1) across sites/scanners and (2) across time. MRI has been employed in numerous different multisite observational and therapeutic CNS studies, including studies on AD [308
], mild cognitive impairment [312
], cerebral vascular disease [308
], and schizophrenia [315
]. The remainder of this section will focus on specific features of the design of multisite CNS studies that were gleaned from the experience of the Alzheimer’s Disease Neuroimaging Initiative (ADNI) [316
]. In multisite studies, enrollment sites are selected on the basis of their ability to recruit and retain subjects meeting specific clinical criteria. Sites are typically not selected on the basis of access to advanced MRI equipment or expertise. Consequently, multisite studies must be performed across a broad array of hardware/software platforms and across sites with variable levels of expertise. State-of-the-art applications may not be available on all systems.
Some basic principles in the design and conduct of multisite CNS MRI studies follow. The main confound to address when different scanners at different sites are used is ensuring that equivalent sequences are being run with equivalent parameters on each scanner to ensure the same contrasts, artifacts, noise, etc. Further, each scanner used in the study should be qualified at baseline before subject enrollment and requalified after any hardware or software upgrades. “Qualified” means using a specific quality assurance (QA) protocol [317
] with a specific phantom to ensure that the scanner is performing adequately. Each site should use the same type of phantom and QA protocol. Ideally, a single scanner should be used at each site for the entire study.
Tesla strength of MRI scanners at different sites may vary. If so, the sequences and specific parameters will need to be adjusted to maintain equivalent T1 and T2 weighting for all scanners across sites. Also, in the case of functional imaging, differences in SNR fluctuations between scanners need to be accounted for in postprocessing before activation across scanners/sites is compared [318
Before start-up of the study, it might be highly useful to conduct a small pilot study of the proposed acquisition protocol on a representative group of systems. Such a pilot study will avoid the unfortunate situation in which incompatibilities between the study protocol and certain platforms are discovered after enrollment has begun. Electronically distributing system-specific protocols to each scanner used in the study is also helpful. Electronic distribution avoids the situation in which protocols are built manually from paper protocols on individual scanners, which dramatically increases the likelihood of protocol errors at individual sites. Imaging sequences in the protocol should be easy to prescribe and use by MRI technologists with a wide range of experience.
Ideally, the study protocol should avoid nonproduct imaging sequences to minimize administrative and regulatory overhead. A central quality control center is recommended. All protocol scans should be checked for protocol compliance, image quality, and medically significant abnormalities. Identification of medically significant abnormalities by the MRI center should be used for subject inclusion and exclusion purposes. However, clinical interpretations of study scans should be the responsibility of the local enrollment site—i.e., medicolegal responsibility should reside locally, not with the research study.
Some means of monitoring scanner performance for site qualification purposes and throughout the study is useful. In the ADNI, a phantom scan is acquired along with each patient study. Measurements from these phantom scans allow identification of scanning errors that elude detection because the relevant information is not recorded electronically with the imaging data. Measurements from phantom scans can be applied retrospectively to correct drifts or discontinuities in the coupled human images, provided certain assumptions are met. Finally, despite detailed attention to standardization of acquisition and quality control, correcting residual abnormalities in the image data after the fact remains important. The types of corrections that may be useful are corrections for gradient nonlinearity, intensity nonuniformity, and drift or discontinuities in scanner calibration.
The preceding comments are largely recommendations derived from the experience of the ADNI. However, these considerations are relevant for any multisite study. To the extent that technical variation is reduced in a data set, the data set becomes a more useful and powerful tool for achieving the ultimate objective, which is the use of imaging as an in vivo surrogate of specific pathologies of interest. The large patient-population capabilities of multisite studies could be invaluable to combat-related TBI and PTSD research. For more information regarding multisite studies, see the ADNI (www.loni.ucla.edu/ADNI
) or the Biomedical Informatics Research Network (www.nbirn.net
) Web sites.