|Additional related information may be found at:|
|Neuropsychopharmacology: The Fifth Generation of Progress|
In Vivo Structural Brain Assessment
In Vivo Structural Brain Assessment
Kelvin O. Lim, Margaret Rosenbloom, and Adolf Pfefferbaum
Over the past few years, technical improvements in the acquisition and processing of structural neuroimaging data, particularly with magnetic resonance imaging (MRI), have provided vivid visual representations of the human brain, both as two-dimensional (2D) slices through the brain, and as three-dimensional (3D) views of the external surface and internal structures of the brain. Computerized tomography (CT) was the first imaging modality to provide in vivo evidence of gross brain morphological abnormalities in schizophrenia (39), with many CT reports of increase in cerebrospinal fluid (CSF)-filled spaces, both centrally (ventricles), and peripherally (sulci) in a variety of psychiatric patients (80). MRI, with its superior tissue contrast for differentiation of white matter from gray matter, and greater flexibility in acquiring, or re-slicing images in orientations best suited for viewing and measuring specific structures, has provided additional capabilities for uncovering both generalized and regionally specific morphometric brain abnormalities associated with schizophrenia, and other mental disorders (46, 88). Patients with psychiatric disorders generally manifest brain abnormalities of a subtle and/or diffuse nature, best characterized by systematic and reliable quantification techniques rather than qualitative assessments.
In addition to characterizing ever more specific brain dysmorphology (the size, tissue composition, and shape of different cortical or subcortical neuroanatomic structures), new advances have also made it possible to use MRI to visualize and quantify the directional coherence of white matter fibers (diffusion tensor imaging, DTI) (9), an application of potential relevance for investigating levels of connectivity and dysconnectivity between different brain regions.
All measures derived from neuroimaging technologies need to be carefully interpreted in the context of other brain and biological features, as well as the clinical and demographic characteristics of the patient, before they can provide insight into the possible etiology, pathophysiology and course of different mental disorders. The goal of this updated chapter is to briefly recapitulate prior information, describe some new advances in neuroimaging and outline their capabilities and limitations and reiterate some of the important technical and methodological issues connected with acquiring, measuring and interpreting these types of data.
ACQUISITION OF MAGNETIC RESONANCE IMAGES
Mechanisms Of Tissue Contrast: Pulse Sequences
One of the strengths of MRI is that investigators can manipulate the amount of contrast between different biological tissues by varying elements of the image acquisition sequence. A highly simplified outline of the principle of MRI is provided here to orient the reader to how these acquisition elements can be varied. More detailed descriptions can be found elsewhere (68). Atomic nuclei with an odd number of nucleons (protons plus neutrons) behave as spinning tops with an electric charge, and thus they produce a magnetic field, oriented along the axis of rotation of the nucleus. Under normal circumstances, the spin axes of these nuclei are randomly oriented so that the sum of all magnetic fields is zero. However, when these nuclei are exposed to a strong external magnetic field, they align either parallel or anti-parallel with the field in an equilibrium state, that is either in a high or low energy state, with a slight preponderance of the high energy state. These aligned atomic nuclei rotate around their axes at a characteristic frequency, also dependent on the strength of the magnetic field. Electromagnetic radiofrequency (RF) pulses, applied at the appropriate frequency perpendicular to the main magnetic field, excite the spinning nuclei, knocking them out of equilibrium. When the radio waves are turned off, the nuclei return to equilibrium. MRI employs appropriately tuned radiofrequency coils to detect the radio signals given off as the nuclei return to equilibrium, and localize them using magnetic gradients in the three orthogonal axes.
Hydrogen atoms (protons) are the most abundant element in biological tissue, primarily in the form of water. Proton MRI thus records signals arising predominantly from free water (i.e., water not biochemically bound in complex molecules), but there is also a measurable but much lesser contribution from protons in fat. Proton density (PD), based on the number of nuclei stimulated, and the relaxation times, T1 and T2, which reflect the chemical environment of the tissue stimulated, each determine in different ways the intensity at which different tissues will appear in an acquired image. T1 is an exponential time constant which represents the time taken for excited nuclei to return to equilibrium after the RF pulse has been turned off. T2 is an exponential time constant describing the time it takes for the excited nuclei to lose signal, mainly due to dephasing in the transverse plane. The time between radiofrequency pulses (TR), and the amount of time after the pulse at which the signal is acquired (echo time, or TE), determine the influence each of these parameters (PD, T1 and T2) have on the acquired image. For example a long TR (2-3 sec), spin echo sequence with a long TE (80 msec) provides a heavily "T2-weighted" image in which CSF remains bright, because CSF has a long T2, while signal from brain parenchyma,which has a shorter T2, has declined, rendering this tisue darker and homogeneous. In contrast, a relatively short TR (200-400 msec) spin echo or inversion recovery sequence with a short TE (20 msec) provides a "T1-weighted" image which optimizes contrast between gray matter and white matter. In T1-weighted images, tissue with shorter T1s are brighter, whereas in T2 weighted images, tissues with longer T2s are brighter. For brain imaging, the choice of T1 vs. T2 weighted images presents a tradeoff between obtaining high contrast between gray matter and white matter or obtaining high contrast between CSF and brain parenchyma. While T1-weighted spin-echo and inversion recovery sequences provide excellent white-gray contrast, and have been used extensively for morphometric studies, they are limited by their poor definition of CSF/skull margins for reliably measuring intracranial volume. Thus, there is no single correct MR brain sequence, but rather a range of options, with the optimal selection determined by the questions being asked.
A major limitation in scanning patients has been the length of time needed to obtain images, resulting in image degradation due to subject motion. A major thrust for development of clinical imaging techniques has been to reduce imaging time, either by reducing TR and TE while incorporating additional pulses to provide contrast, or by acquiring more data after each excitation pulse. Techniques that acquire the same amount of data in less time produce image contrasts similar, but not identical, to those of traditional spin-echo sequences. Sequences that acquire all of the data needed for reconstructing an image after a single excitation pulse (single shot techniques) include Echo Planar Imaging (EPI) (53) and spiral imaging (61) techniques. While fast and convenient, these images often suffer from lower signal-to-noise ratio (SNR) and susceptibility induced distortions. Work is constantly under way by the manufacturers and research groups to produce further enhancements and overcome limitations. Approaches that minimize the influence of susceptibility artifacts of EPI based systems by acquiring only part of the data after each excitation include Rapid Acquisition with Relaxation Enhancement (RARE) and Fast Spin Echo (FSE) sequences.
Sources of contrast other than that based on manipulation of T1, T2, and proton density have been exploited to extend the information available through MRI. One of these is based on magnetization transfer, which exploits the difference between free and restricted (typically in myelin) hydrogen molecules. Off-resonance RF pulses are used to selectively saturate the restricted pool, inducing an exchange of magnetization between free and restricted hydrogen protons and promoting T1 relaxation. The resulting contrast is believed to reflect the structural integrity of the tissue being imaged (27, 58) particularly white matter.
Another source of contrast is based on the self-diffusion or the random Brownian movement of water molecules. When diffusion is unrestricted, e.g. as in a glass of water, the movement of water molecules has no directional preference and is termed isotropic. However, when the motion is restricted, e.g., by fibers such as constitute the white matter tracts, diffusion is greater parallel to the fiber than perpendicular to it, and is termed anisotropic. Anisotropic diffusion can be imaged, and by sequentially applying magnetic gradients in six non-collinear directions, its directionality and magnitude can be quantified. This form of imaging has been called diffusion tensor imaging (DTI) (9). Subsequent analysis applies a mathematical construct, a tensor, to describe the multidimensional vector systems and quantify the diffusion. Vector measures have been used to map white matter fiber tracts in the brain. Scalar measures of the diffusion, such as fractional anisotropy and relative anisotropy, provide an overall index of tissue anisotropy (52). This approach has recently been used to reveal possibly compromised white matter tracts in patients with schizophrenia (13, 50).
Two-dimensional Multi-slice and Three-dimensional Imaging
Two-dimensional (2D) images are acquired in a defined orthogonal plane, typically in axial, sagittal and coronal orientations (Figure 1). Image orientation is defined by use of selective RF pulse excitations and the appropriate magnetic gradients in the three orthogonal axes. Each orientation provides a different view of the brain with optimal visualization of different structures. For example, mid-sagittal images provide clear delineation of the prefrontal cortex and the corpus callosum, coronal images provide views of the hippocampus and limbic structures over several sections or slices and axial images provide a good view of basal ganglia structures such as the putamen, globus pallidus, caudate and substantia nigra, as well as the lateral ventricular system. One limitation of 2D image acquisition is that only selected slices, of specified thickness, are excited and imaged and, unless these slices have been collected using standardized internal neuroanatomic landmarks, they may not be directly comparable across subjects. Comparability is enhanced by orienting each slice acquisition relative to a specific anatomic plane such as that passing through the anterior and posterior commissures, perpendicular to the sagittal plane (95).
Three-dimensional (3D) volume acquisition protocols encompass the entire brain and can be reformatted, post acquisition, into any plane subject to the limitations imposed by the resolution of the original data. This post-acquisition flexibility in defining 3D planes for analysis can reduce measurement errors related to head misalignment between subjects, or within a subject in a longitudinal study. 3D images are acquired by exciting a broad volume of tissue with a non-slice selective RF pulse (5). Additional magnetic gradients are then used to spatially encode within the excited region. After the raw image data have been collected, a Fourier transform decodes slice positions from the entire data set. These slice data are further reconstructed with a 2D Fourier transform into the final image. Limitations in post-scan analysis of 3D data sets are discussed in more detail in the section on image analysis below. Among the 3D acquisition protocols which are widely used in psychiatric neuroimaging are spoiled gradient recall acquisition (SPGR) and magnetization prepared rapid gradient echo (MPRAGE) (63) each of which are T1-weighted, provide reasonable gray/white differentiation, and can sample the entire brain with 1.5mm or thinner slices in 10 minutes or less.
Image Resolution, Signal-To-Noise-Ratio, and Slice Thickness
An image is composed of a grid (usually 256 x 256) of 2D picture elements (pixels). The resolution of the image is affected both by in-plane resolution and by slice thickness. An in-plane resolution of 1 mm means that each pixel in the image matrix represents 1 mm2. However if the slice is as thick as 5 mm, once standard for 2D MRI protocols, each image pixel actually represent a voxel volume of 5 mm3. The larger the voxel, the greater the likelihood it will be "partially volumed", i.e., contain some combination of CSF, white matter or gray matter rather than being homogenous for brain tissue type. To reduce partial voluming, high in-plane resolution and thin slices are desired. However, technical factors such as gradient strength and the amount of RF energy used to excite the object, limit the minimum thickness effectively available to approximately 1-2 mm. In addition the SNR per voxel will be lower because there is less material producing the signal in a thin slice. Thus, while the resolution may be "finer" there will be more noise and less signal per pixel. Thinner slices may also require an increased number of acquisitions and thus longer scanning time to encompass an equivalent volume of the brain, with the attendant disadvantages of requiring subjects to remain still in the scanner.
MR images are susceptible to various types of artifact, some of which can be minimized during image acquisition and others which can only be remedied during data analysis.
Hardware Induced Artifact
The RF coil's sensitivity is not homogeneous across all dimensions of the imaging plane which often introduces a low frequency gradient of signal intensity across a given image. While the human visual system can maintain contrast detection in the face of this artifact, most automated thresholding techniques assume a constant baseline level and require the RF inhomogeneity to be corrected before the image can be processed. A variety of solutions have been proposed to this problem, including filtering (51) bi-feature thresholding (44), and line by line inversion of the slope of the pixel values within each slice (17) and adaptive filtering (107).
Effects Of CSF And Blood Pulsation, And Subject Movement
Physiological sources of artifact include pulsation of blood and CSF through regions of the brain being imaged. The limbic system is particularly vulnerable to this artifact because of its vascularization pattern and relationship to the ventricular system. Cardiac-gating and flow compensated pulse sequences help reduce these sources of movement artifact, although not without cost. Because more gradients are used with flow compensation, the minimum echo time is lengthened and the maximum number of slices that can be acquired is reduced. Voluntary or involuntary gross head movements are another source of artifact to which 3D, functional MRI and DTI are particularly susceptible. Head restraints or bitebars (41, 60) provide one solution A complementary approach, adopted by some research groups for structural imaging, involves administering a sedative hypnotic, adequate to calm the subject or induce sleep.
Need For Phantoms
The accuracy and reproducibility of image data is directly related to the stability of the imaging hardware, especially the magnetic field gradient systems, the main magnetic field, and the RF pulse system. Routine imaging of a quality assurance phantom permits monitoring of the scanner's stability over time, stability between hardware upgrades, and establishment of standardized techniques for multi-center studies (101).
A recurring theme in any discussion of MRI data acquisition for clinical studies is the constant tradeoff an investigator must make between resolution of the image, its SNR, and time of the subject in the scanner. Even with cooperative and relaxed subjects, involuntary head movements during a lengthy scan session can be a problem. Optimal image acquisition parameters for specific studies need to be determined empirically and will vary depending on the structure being imaged, scanner hardware, and the clinical characteristics of the subjects being studied. The high cost of MRI scanning time, along with limitations in patient tolerance, combine to favor short acquisition sessions. However, reductions in scanning time may be at the cost of poorer image resolution and contrast, or sampling less of the brain. A more detailed technical treatment of pulse design issues and their relative strengths and weaknesses is available elsewhere (e.g., 85).
ANALYSIS OF MAGNETIC RESONANCE BRAIN IMAGES
Automated computerized analysis of many elements or characteristics of an image is becoming increasingly feasible and accessible to investigators. These approaches offer considerable advantages of flexibility and efficiency, though even automated analysis approaches may not overcome all limitations of image resolution, and distortions due to artifact in data collection. Most analysis approaches combine both automated and interactive elements (e.g., 24). Clinical reading or qualitative scoring of MRI brain data, still widely used by neuroradiologists, serves as a useful complement to quantification, particularly for assessment of image features that until recently have defied reliable quantification such as white matter lucencies or gross "abnormalities" in brain morphology. However recent advances in characterizing tissue quality (e.g., 96) and shape (e.g., 10, 19) of defined structures are making even these abnormalities amenable to quantification.
Resolution is limited by parameters of the data collection protocol, including in-plane resolution and slice thickness. Generally 2D data are analyzed as acquired, with a known in-plane resolution and slice thickness. In contrast, 3D data sets offer the opportunity for reformatting images in user-defined planes in any orientation and slice thickness. This provides greater flexibility in optimizing the view of specific structures. However, re-slicing is bound, not only by the hardware and software constraints of the acquisition parameters, but also by constraints of the object space ( i.e., the defined plane and thickness) of the original images. The effective thickness of a user-defined slice will depend on the orientation of the slice in the 3D volume. For example, with an isotropic data set, (i.e., each voxel is a true cube with side length = L) where resolution in the x, y, and z axes is equal, the maximum distance between opposite corners of the cube will be sqrt(3)*L. Hence a slice reconstructed normal to a line connecting these two corners would have an effective slice sampling of sqrt(3)*L. This orientation-dependent effective slice-thickness needs to be kept in mind when analyzing such data. This problem is compounded further when the data set is anisotropic, as are most.
Gross Morphology/Issues Of Reliability And Validity
MRI provides great flexibility in acquiring a number of different visual representations of the human brain, each designed to enhance contrast between different tissue types, highlight abnormalities within a tissue type, or provide an optimal view of a specific lobar region or subcortical neuroanatomic structure. Techniques for morphometric analysis rely to varying degrees on user interaction with a visual representation of the brain image, displayed on film or a computer console, and automated algorithms applied to the matrix of numerical values representing each pixel in the image. Validity must be established for both interactive and automated approaches, though in the absence of satisfactory in vivo validation models, automated approaches are frequently validated against the "gold standard" of a human expert, by establishing consistency with known effects such as changes with age, by comparisons with pathological samples, or by use of phantoms simulating different tissue types (1). Given the complexity and individual variability of each brain, issues of reliability acquire a particular salience when human input is involved, but should also be documented for automated elements of any analysis procedure (e.g., 8, 105)
The size of structures of the brain can be estimated using linear, area, or volumetric measurements. In the early CT literature, planimetric (area) measures were most commonly used and applied to one or two "best view" slices. This approach was accessible to investigators without computerized image processing facilities, but was vulnerable to sampling error, and sacrificed volumetric information. With the advent of MR, and increasing availability of computerized analysis, volumetric analysis has become the norm. A volumetric approach, ideally encompassing the entire structure in each hemisphere, is particularly important for studying left-right asymmetry of structures that are not evenly aligned relative to the plane of view on a single slice of a given image (112). This approach can compensate for any bias in selection of imaging plane, or apparent asymmetries due to head placement.
CSF, Gray Matter and White Matter
Before brain tissue can be segmented, its extent must be defined. In vivo estimates of brain size are usually based on estimates of the total volume of brain tissue and CSF, including both ventricular and subarachnoid fluid. Non-brain material must be identified or removed from the image before the brain as a whole can be measured and its external features identified. Correct identification of the outer brain/dural boundary involves certain assumptions and estimations and is complicated by the fact that bone can easily be confused with CSF and tissue in many MRI acquisition sequences, especially T1 sequences designed to enhance contrast between white matter and gray matter. Such ambiguities may underlie the recently observed variation in estimated brain size between a number of MRI studies (71). One solution to the problem is to use late echo images in which CSF, with its long T2, can be sharply differentiated from bone, for identifying subarachnoid fluid margins and then transfer information about these boundary locations to images obtained with sequence parameters on which CSF/tissue and white/gray matter contrasts are optimized (51). Although brain tissue volume is the principal object of study in the effort to identify neuronal substrates of dysfunction, the size of CSF-filled spaces relative to the size of the intracranial vault can also be instructive for interpretation regarding degeneration of tissue and progression of disease.
Tissue segmentation routines to differentiate CSF from tissue, and subdivide tissue into gray matter and white matter all have limitations because there is no discrete image intensity signature, even across echo times, for each tissue type. Segmentation is also complicated both by the presence of areas of white matter hyperintensity (WMHI), and by the fact that individual voxels may be partially volumed and thus contain mixtures of CSF and gray matter or gray matter and white matter. Approaches to this analysis include edge detection, application of various forms of filters, and fuzzy clustering, each of which can be implemented with varying degrees of operator interaction. Edge detection, for example, of the outer margin of the brain, the ventricles, white matter tracts, or specific subcortical structures such as the hippocampus or striatal structures (89) can be done by a human tracing outlines on a computer display or by using algorithms which identify pixels whose value differs sufficiently from neighboring pixels to qualify them as a boundary pixels (7). Such automated edge detection works well when contrast is high and structures to be outlined are of a relatively simple shape. Operator input frequently supplements automated outlining.
Thresholding involves classifying pixels according to image intensity as representative of CSF or different tissue types. Some approaches take advantage of the different contrasts obtained by different pulse sequences or echo times, either combining data from T1, T2 and proton density weighted images (e.g., 1) or subtracting one image from another to enhance CSF-tissue contrast (32, 51). More recently, "fuzzy" cluster or classifier approaches, which adopt a probabilistic model towards each voxel and thus are better able to account for partial voluming effects have been found to have high reliability, accuracy and validity (33, 82). Automated assessments may employ discriminant functions based on ranges of pixel values representing known tissue and fluid samples (e.g., 14, 17).
One interactive approach requires an operator to adjust a 2-bit image until a pixel value is identified which correctly differentiates CSF from tissue (51). Another takes advantage of the fact that while losses in signal differentiation, for whatever cause, make automated segmentation more difficult, the human visual system, which is quite robust to local intensity changes, can often compensate for them—people can see the gray-white tissue differences even when the segmentation routines fail. In this situation, stereological point counting methods (11, 28, 42) might be superior. This approach involves the random placement of a grid with sufficient resolution (in two or three dimensions) over the structure of interest and counting the points overlying the region or structure of interest. The only requirements are that the grid completely encompass the region or structure, that it be placed on the structure randomly, and that an adequate number of points (150-200) be counted on an adequate number of slices (~10). The approach can be very efficient and is statistically sound. The stereological point counting approach has the additional advantage of providing a coefficient of error of the measurement of the volume of the structure of interest which estimates the magnitude of difference that can be determined between subject groups.
Models for tissue segmentation which work well on healthy young brains may fail when applied to aging subjects or populations with greater tissue pathologies, particularly of the white matter (e.g., 73). Furthermore, assessments of subcortical gray matter volume using thresholding algorithms can be affected by variations in tissue iron content which shortens T2, resulting in signal loss. Iron content-related changes in T2 occur with normal aging and in basal ganglia disorders, particularly in the substantia nigra and globus pallidus (78). The misclassification of WMHI as gray matter can be handled by visual checking and manual correction (e.g. 75) in samples where this occurs infrequently. Clinical populations where focal abnormalities such as gliosis, edema, demyelination, and ischemia, processes that alter the properties of tissue water and produce changes in white and gray matter signal intensity, are common require new approaches to tissue segmentation. Magnetization transfer ratio is one such approach, specifically designed to characterize white matter pathology (96, 103).
Structures And Regions Of Interest For Measurement
As in image acquisition, where tradeoffs need to be made between time in scanner and optimizing image resolution, tissue contrast, and range of structures visualized, so in image analysis there are tradeoffs to be made between labor intensive, interactive delineation of boundaries, and automated approaches, whether using geometrically based boundary drawing algorithms (e.g., 111), parcellation against a standardized set of coordinates (e.g., 4), or template matching (98). As research laboratories collect ever increasing numbers of brains for analysis, automated approaches which can be implemented quickly and with less ongoing neuroanatomic expertise become very attractive. The jury is still out on whether the advantages of speed and reliability of such systems offset inaccuracies introduced by the deformations involved in a "one size fits all" measurement approach. Approaches combining automation with expert supervision attempt to get the best of both worlds.
One approach to automated measurement of major brain regions uses an MRI-derived anatomic atlas, based on a stereotactic coordinate system (95). Individual 3D brains are linearly transformed to fit neuroanatomic anchor points from the atlas and volumes of ROIs are computed by summing brain voxels coinciding with the atlas boxes assigned to each ROI (e.g., 4). This approach has been extended by thresholding brains into CSF, white, and gray matter to delineate subcortical gray matter structures as well as overall cerebrum, cerebellum and ventricular volumes (40) and validated against interactively defined gold standards and by demonstrating predicted gender, age and group effects. For a more comprehensive treatment the variety of brain atlases available, and different approaches to deformations the reader is referred to a recent specialized publication (102).
While much can be inferred from global measures of overall brain volume and volume of different tissue compartments, specific regions particularly the frontal and temporal lobes as well as limbic and striatal structures are believed to be of particular interest in characterizing the pathophysiology of psychiatric disorders (12). New capabilities for 3D display of brain images have greatly expanded options for delineating brain regions according to sulcal and gyral landmarks, but also can confound the investigator with the complexity and variability of the human brain. For example, while the central sulcus provides the boundary separating frontal lobes from parietal lobes, its configuration varies between individuals and between each cerebral hemisphere. One of the earliest MRI estimates of frontal lobe volume was based on manual tracing of a single midsaggital slice (2), yielding a single measure representing a fraction of the entire structure. Now investigators are combining surface sulcal landmarks (79) with multiple, orthogonal views of brain tissue to identify and trace boundaries of specific cortical gyri (11, 108), although these are not the same as functionally homogenous cytoarchitectonic brain regions. Among the subregions defined and segmented into white and gray matter are superior, medial, inferior, orbital, insular and cingulate regions. This painstaking approach requires interactive neuroanatomic judgments, considerably speeded by visual enhancement of sulcal boundaries, and simultaneous display of internal images in different orthogonal planes. A recent application of this technique to the prefrontal cortex in schizophrenia found a gray matter deficit only in the inferior region (11).
Rules for defining boundaries for temporal lobe as a whole, and its subsections such as the superior temporal gyrus (STG) and planum temporale have been constrained in many cases by the resolution and plane of view available to the investigator, and have not been comparable across laboratories. Increasing availability of high resolution 3D acquisition protocols, and interactive 3D image processing systems will enable closer approximation to established anatomical criteria, and hopefully the adoption of standardized criteria. Absence of standard definitions for mesial temporal lobe structures, particularly hippocampus and amygdala, has been particularly evident in MRI studies of these regions in patients with schizophrenia (for review, see 35). Meta analysis, which provides greater power to detect effects submerged in varying sources of error in individual studies, indicates that mesial temporal structures are reduced in volume in schizophrenia, whether or not the amygdala is included in the measure (66). The intensely demanding requirements for interactive delineation of hippocampal structures, as well as the error inherent in subjective assessment of structure boundaries motivated development of an automated approach, that measures not only volume (31) but also characterizes shape (30). Application of multi-dimensional transformations based on pattern matching has identified specific shape deformations of the hippocampus in patients with schizophrenia localized to regions of this structure that send projections to the frontal cortex (19). Such analysis was not only more sensitive in differentiating patients from controls than simple volumetric measurements, but also identified an abnormality consistent with the hypothesized disconnection of hippocampal-prefrontal circuitry.
The development and application of laboratory-specific boundaries, criteria, and procedures for defining neuroanatomic regions will lead to inconsistencies rather than convergence in the literature.
Shape and Complex Morphological Features.
New analyses incorporating advances in physics and mathematics are now enabling examination of both focal and global characteristics of a structure that are independent of variations in size, and can be performed with varying degrees of automaticity. Until recently, a 3D representation of brain structures could only be visualized in the investigator's imagination. Now 3D displays of the external surfaces, with ability to slice through selected planes are routinely available in most brain imaging laboratories. Creation of 3D images employs a range of rapidly evolving computer graphics software and hardware methods (23), surface and volume rendering techniques. While visually compelling, rendered images may contain distortions introduced by the viewing model and image preprocessing that must be taken into account before quantitative analysis is made of the rendered image.
Neurosurgeons and neurologists have lead the way in applying these techniques for localizing lesions and vasculature before neurosurgery (e.g., 20), developing neuroanatomic teaching tools enabling students to explore the structure and function of the brain in virtual reality (e.g., 99), characterizing the normal surface anatomy of the brain (104), and identifying specific neuroanatomic regions based on specific sulcal and gyral landmarks (79). Fluid deformations of one exterior brain surface against another (template) followed by encoding and mapping of probability of displacement are being used to assess individual and group differences in the immensely complex and variable gyral and sulcal topography (98). These probability statistics expand on earlier concepts of anatomic atlases by simultaneously characterizing both locational and geometric variation and open up a new dimension in characterizing and quantifying regional abnormality relative to age and gender norms. Other structures for which shape analyses have uncovered specific dysmorphologies in schizophrenia include the hippocampus (19) and the corpus callosum (16, 97).
go to Figure
go to Figure 2
With Other In Vivo Neuroimages
The relatively poor in-plane resolution and wide slices of images from functional imaging techniques such as positron emission tomography (PET) and proton or phosphorus magnetic spectroscopic imaging (MRSI) limits anatomic localization. Structural information from MR images is now routinely used for localizing and displaying high intensity areas of activation derived from functional imaging modalities including PET (67, 91) and functional MRI (fMRI) (15). Indeed, this localization of functional data is important not only for more precisely characterizing the structures involved, but also to take regional or group differences in CSF, gray and white matter composition of the partially volumed functional voxels into account in interpreting results. Since voxels obtained in MRS or PET tend to be larger than those now available with MRI, co-registration of MR and PET or MRS images allows voxel composition information from structural images to be applied to functional images. This approach has been used for PET (59, 64), and recently also for MRSI (34, 49, 75). Such corrections are computationally challenging, but eventually will allow much greater precision in assessing functional activity.
The simplest case of co-registration is when both data sets have been collected on the same imaging hardware, without moving the subject, as is typically the case with fMRI and MRSI scanning (Figure 3). For PET scanning, differences in hardware and time of scan are unavoidable. Several different approaches to image acquisition to facilitate subsequent matching of acquired PET and MR images have been developed. These include use of external landmarks, or fiducials, consistently placed and of a size and property to be adequately imaged by both modalities (22), and head holders to stabilize and standardize head position between scanners (41). Varying levels of sophistication have been applied to compute a transformation matrix between two sets of images, including iterative matching of brain surfaces (48, 70) projections (93) or overall brain tissue (25, 110). The automated image registration (AIR) (110) and statistical parametric mapping (SPM) (25) systems are now the most commonly used. They have recently been cross-validated and their reliability established (43)
A particular application for merging of data sets within an imaging modality is the generation of averaged brains (3, 109). This approach involves aligning and matching successive images relative to stereotactic coordinates (21) and provides a useful visual summary of group differences. Regions of statistically significant differences can also be identified, but should be verified by a separate case by case quantification of the structure or region in question (72). Image averaging works better for assessing subcortical and midline structures, less affected by deformations necessary to standardize brains, than cortical structures.
go to Figure 4
INTERPRETATION OF MAGNETIC RESONANCE BRAIN IMAGES
Neuroimaging findings in some mental disorders can be elusive. The true effect is small, and brain morphology is strongly affected by several variables which are relatively independent of disease, such as age, sex and somatic size, as well as those which might be associated with mental disorder, e.g., alcohol consumption, socio-economic status and handedness. These variables must be taken into consideration and appropriate controls implemented before inferences about the pathophysiology of various mental disorders can be properly drawn.
Selection of Appropriate
When CT was the prevailing technology, medical patients with "normal" scans were frequently used as controls in imaging studies of psychiatric patients - a practice motivated, to some extent, by reluctance to expose healthy subjects to unnecessary radiation. Whether or not this practice contributed to observed group differences was never fully resolved (81, 92). With MRI scanning, this radiation concern is not relevant. However, the prospective recruitment of control samples for any neuroimaging studies of psychiatric disorder requires attention not only to excluding cases with psychiatric disorders (90), but also to including subjects who match the parental socioeconomic status, educational and ethnic status of the patients (84). In addition to comparisons with normative data, comparisons between pathologic groups are also valuable, especially between different diagnostic groups with documented abnormalities in common anatomic region(s). For example, comparisons between patients with temporal lobe epilepsy and schizophrenia are appropriate for studies investigating the role of limbic structures in these disorders (55).
Normal Developmental and Aging Effects
Schizophrenia is typically a disease with onset in late adolescence or early adulthood, a period during which cortical gray matter undergoes significant reduction, but cortical white continues to increase in volume (74). Its course coincides with a period of stability of white matter volume, but gradual reduction in gray matter and increase in CSF (37, 65, 74, 83) (Figure 5). These normal developmental and aging effects need to be taken into account when interpreting observed brain dysmorphology in schizophrenia. Studies of Alzheimer's disease (AD), Parkinson's disease (PD) and other diseases which typically make their appearance after middle age, likewise need to take into account not only normal age-related changes occurring at this point in the life span, but also the greater variability in brain tissue volumes found among the healthy elderly than the healthy young (37, 38).
Normal Variations In Somatic Size
Given the existence of a wide range of body and head sizes among the population, it is important to demonstrate that any size difference of a specific brain structure is independent of such non-disease-relevant differences. In general, large people will have larger heads and larger brains than small people, although such relationships have been easier to demonstrate across gender and between racial groups than within them (71). In most research contexts, this variability constitutes a form of noise which investigators need to minimize in order to compare the volume of a particular brain region or structure between individuals and groups. Thus head size correction seems particularly appropriate when gender differences in brain morphology are under investigation (18, 45).
One approach for head-size correction has been use of a ratio, such as the ventricle brain ratio (VBR), which was developed to provide an index of ventricular enlargement relative to the size of the total brain. Other correction approaches have included proportions, where the size of a structure, or the amount of tissue or CSF in a ROI, is divided by an estimate of the intracranial volume (e.g., 37); or regression analysis which partials out the components of quantitative brain measures attributable to headsize variation (e.g., 74).
Concern that the ratio or proportion approaches to head size correction could introduce error or obscure the complexity of underlying relationships has lead some investigators to warn against their use, and advocate regression or multifactorial approaches, as necessary, to parse out contribution of non-specific variables (6). Others (57) have demonstrated that a regression approach to head-size correction removes irrelevant true-score variance which reduces reliability and improves the correlation with validity criteria such as age and diagnostic status. Linear regression of specific ROIs against estimated head-size in a sample of normal control subjects highlighted the fact that different parts of the brain bear different relationships to overall brain size and suggests that ROI-specific correction factors should be applied (57).
While head-size correction is appropriate in some contexts, in others, such as developmental studies (74), it may not be. For such studies, absolute volumes of particular brain regions rather than head-size corrected values should be used. Because investigators have described reduced intracranial volume in patients with schizophrenia (69), the issue has become more complicated. If schizophrenia affects the brain before it has completed its growth, leading to diminished brain size and as a result diminished skull size, the correction of all structural brain measures for intracranial volume could obscure pathologically relevant information. Regardless of possible disease-related reduction in head size, however, an investigator asking whether particular brain structures are disproportionately reduced, must examine and take into account differences in intracranial volume.
Widespread Tissue Deficits
Generalized morphological differences in the brains of psychiatric patients compared to controls should be taken into account before assessing smaller structures. For instance, schizophrenia appears to be characterized by generalized ventricular and cortical sulcal enlargement as well as a widespread loss of cortical gray matter (88). Therefore, it is necessary to consider whether loss of tissue from the target region exceeds that found from adjacent or control regions before concluding that a specific brain structure or region is uniquely affected by the disease (e.g., 112). Using such an analysis, recent studies have indeed identified certain regions of frontal and parieto-temporal cortices (86, 94) as particularly affected in schizophrenia.
As neuroimaging matures and becomes a standard element of psychiatric clinical research, longitudinal studies of change over time will increasingly replace the inferences made from cross-sectional studies. Direct longitudinal comparisons over time are hampered by major advances in technology (e.g. replacement of CT with MRI) as well as system upgrades within a technology. However, the advantages of direct longitudinal evaluation of the progression of a disease over inference from cross sectional studies motivates the development and application of procedures both to standardize collection parameters over time and to estimate and correct for the error introduced by uncontrolled factors.
Within psychiatry, longitudinal neuroimaging studies are beginning to show promise for studying the progression of the disease (29, 56, 76), ideally with a matched control group also followed for comparable intervals. Neuroimaging studies of multiple sclerosis, where the waxing and waning of white matter lesions are regularly monitored provide some important pointers for longitudinal studies of psychiatric patients (62). Studies of multiple sclerosis were among the first to alert investigators to the influence of measurement error due to misalignment between scans over time on measures of real change (26). Other subtle sources of error between tests identified by MS researchers include variation in scanner sensitivity (106) and voxel dimensions (47). The use of phantom standards in all scans greatly facilitates application of post hoc adjustments (101).
Measures In Face Of Changing Technology
As neuroimaging methods advance and new imaging techniques are used, it is necessary to understand how morphometric methods are affected by differences in imaging methods. A major transition in the past 15 years has been the replacement of CT scanning with MRI scanning. While MRI adds much information, particularly from its enhanced soft-tissue resolution, to that provided by CT, it also provides similar information on CSF-filled spaces, particularly ventricular enlargement, which has been a hallmark of CT studies. Thus MRI values can potentially offer continuity for longitudinal studies in which a transition was made from CT to MRI. A comparison of CT and MRI-derived estimates of ventricular and sulcal CSF volume obtained within two weeks of each other in patients meeting RDC criteria for alcohol dependence (77) found that while CT and MRI estimates of absolute ventricular volume were equivalent, CT considerably underestimated cortical sulcal volume relative to MRI. This underestimation could be due to poorer resolution and spectral shift artifact on CT. Thus longitudinal studies of subjects initially scanned with CT and subsequently with MRI may, with suitable caution, assess changes in ventricular volume, but should not compare sulcal volume.
The increasing replacement of conventional 2D image acquisition sequences with fast 2D and 3D acquisition sequences makes issues of comparability across acquisition sequence within MRI extremely relevant. While comparisons between MRI sequences for sensitivity to detecting various types of lesions are common (e.g., 100) few comparisons between regional volumes of CSF, white and gray matter obtained by different sequences have been published. One pilot study, tested the comparability of CSF, GM, and WM measures made from 3D SPGR images to those acquired using a T2-weighted axial spin-echo protocol (54) and found that although there was a high degree of correlation between the CSF, white and gray matter values from both sets of images, there was relatively greater white matter and less CSF and gray matter in SPGR compared to spin echo images, suggesting a systematic difference between the methods. Thus longitudinal comparisons of data collected using different acquisition methods should be made with caution since the differences in tissue contrast, signal-to-noise ratios, and slice profiles will affect the results of morphometric measures.
Even when software protocols are identical, differences in head positioning between scans can contribute measurement error in studies involving repeat scans (26), particularly when 2D, rather than 3D images are being compared. A longitudinal CT study on patients with AD identified apparent intra-individual change in total cranial volume between first and second scans ranging from -60.5 to 89.5 cc, which correlated significantly, but differentially with several of the ROI change scores. Since total cranial volume (CSF plus tissue) should not change between scans in adults, this apparent change was used to derive a specific correction factor for each ROI change (87). This procedure has since been applied to longitudinal studies of patients with alcoholism (76) and schizophrenia (56). Other approaches to correcting for differences in head positioning between scans include realignment of scans in three axes and reslicing along the AC-PC line (29). These efforts to correct measurement error due to differences in head placement have only recently been applied in longitudinal studies, and the lack of such corrections in earlier longitudinal studies of brain morphological change in patients with schizophrenia may account for inconsistent results.
Other sources of error include unreliability in measurement techniques (36) as well as changing approaches to measurement. Problems limiting the value of earlier neuroimaging follow-up studies can now be substantially reduced by applying the same measurement approach to baseline as well as follow-up data sets, assessing inter and intra-rater reliabilities, and retesting control groups along with patient groups.
The superior resolution and flexibility of MR imaging has enabled the examination of more detailed and specific anatomical questions in psychiatric disorders. New developments in acquisition and image processing software continue to expand the possibilities available to investigators, not only for visual presentation, but also for quantitative analyses. While there continue to be important methodological and technical issues to attend to, investigators are becoming increasingly sophisticated in using appropriate techniques, and adopting strategies to ensure the reliability and validity of the data being produced. Standardization of techniques between laboratories is becoming increasingly viable though the dissemination of image analysis software and posting of brain atlases in web sites. It continues to be important that data are collected and analyzed in such a way that specific hypotheses about particular regions of the brain can be tested within the context of the brain as a whole.