Quantitative Neuromorphometry Using Magnetic Resonance Histology

Abstract

Magnetic resonance imaging (MRI), now common in the clinical domain, has been adapted for use by the neuropathologist by increasing the spatial resolution over 100,000-times what is common in human clinical imaging. This increase in spatial resolution has been accomplished through a variety of technical advances—higher magnetic fields, more sensitive receivers, and clever encoding methods. Magnetic resonance histology (MRH), i.e. the application of MRI to study tissue specimens, now makes three-dimensional imaging of the fixed brain in the cranium routine. Active staining (perfusion fixation with a paramagnetic contrast agent) has allowed us to reduce the scan time by more than 8-times over earlier methods. The result is a three-dimensional isotropic image array that can be viewed along any direction without loss of spatial resolution. Homologous slices can be chosen interactively. Since the tissue is still fully hydrated in the cranium, tissue shrinkage and distortion are virtually eliminated. Volume measurements of neural structures can be made with a high degree of precision and accuracy. MRH will not replace more traditional methods, but it promises enormous value in choosing particular areas and times for more traditional sectioning and assessment.

Introduction

Magnetic resonance imaging (MRI) has revolutionized modern clinical medicine. Paul Lauterbur and Peter Mansfield, who shared the 2003 Nobel Prize in Physiology and Medicine for their invention of MRI, both recognized the potential for the application of MRI beyond the clinical domain. The concluding sentence of Lauterbur’s paper notes that “Zeugmatographic (imaging) techniques should find many useful application in studies of the internal structures, states and compositions of microscopic objects” (Lauterbur, 1973). Peter Mansfield recognized in one of his early papers “Diffraction in microscopy in solids and liquids by NMR” that MRI could be performed at spatial resolution far higher than the limit imposed by conventional diffraction (Mansfield and Grannell, 1975). Magnetic resonance microscopy (MRM) was introduced to practice in 1986 with the first laboratory demonstration of MRM (Aguayo et al., 1986; Eccles and Callaghan, 1986; Johnson et al., 1986).

Over the last 25 years, steady technical progress has been made in magnetic resonance microscopy with applications in a wide range of animal models, drug discovery, and safety assessment. The demonstration of magnetic resonance histology (MRH) in 1993 (Johnson et al., 1993) introduced a particularly novel application for pathologists. While there have been a number of reports of MRH in neurotoxicology (Lester et al., 1999; Lester et al., 2000; Maronpot et al., 2004; Morgan et al. 2004; Sills et al., 2004), broad use of the technology has been limited by spatial resolution, complexity, availability, and cost. This paper will provide an update on efforts to remove these barriers, thus making MRH more widely available to the pathology community.

Materials and Methods

All imaging studies were conducted at the Duke Center for In Vivo Microscopy, an NIH/National Center for Research Resources (NCRR) National Biomedical Technology Research Center, under guidelines approved by the Duke Institutional Animal Care and Use Committee. Tissues were prepared using the active staining protocol described more fully elsewhere (Johnson et al., 2002). The same perfusion technique is used for both the mouse and rat. Animals are brought to a surgical plane of anesthesia. The thoracic cavity is opened and a small catheter is inserted into the left ventricle, and the right atrium is then cut for drainage. The animal is perfused initially with a mixture of saline and ProHance (10:1 saline:ProHance) (Gadoteridol, Bracco Diagnostics, Inc., Princeton, NJ), followed by a mixture of ProHance in 10% neutral buffered formalin (10:1 formalin:ProHance by volume). The head is removed from the body and further fixed for at least 24 hours by immersion in neutral buffered formalin before imaging. On the day of imaging, excess skin, muscle, and lower jaw are trimmed from the head with the brain still in the cranium, so it fits tightly in an acrylic holder filled with fomblin (Ausimont USA, Inc., Thorofare, NJ) (a perfluorocarbon used to limit magnetic susceptibility artifacts); this holder has been designed explicitly for MR histology.

Data shown here were acquired on three systems operating at 9.4T, 7.0T, and 1.0T. The 9.4 T system is an 89-mm vertical bore (50-mm clear bore) superconducting magnet with Resonance Research gradients (Resonance Research, Inc., Billerica, MA) capable of reaching 2000 mT/m. The 7.0T system is a 210-mm bore (90-mm clear bore) horizontal superconducting magnet with Resonance Research gradients achieving 740 mT/m. Both the 9.4 and 7.0 T systems are interfaced to General Electric (GE Healthcare, Milwaukee, WI) imaging consoles that have been modified for MR microscopy. The 1.0 T magnet represents a more economical new approach using a permanent magnet. The self-shielded permanent magnet requires very little site preparation, uses no cryogens, and is significantly less expensive. The computer console has been designed for simple imaging protocols and higher throughput required in screening studies.

A range of imaging software has been used for this work. For the highest-resolution studies from the 7.0T and 9.4T systems, data were reconstructed on specialized offline reconstruction engines that accommodate very large image arrays (to 4096³) (Johnson et al., 2007). Image-processing pipelines have been developed on two dedicated workstations for image segmentation using tools implemented in Advanced Normalization Tools (ANTS) (http://www.picsl.upenn.edu/ANTS/) and the National Library of Medicine’s Insight Segmentation and Registration Toolkit (ITK) (www.itk.org); diffusion calculation using DTIStudio (https://www.mristudio.org/); and visualization using a combination of ImageJ (http://rsbweb.nih.gov/ij/), Avizo (VSG, Visualization Sciences Group, Burlington MA) (http://www.vsg3d.com), and Vitrea (Vital Images, Inc., Minnetonka, MN) (http://www.vitalimages.com/home.aspx).

Results

Active staining allows one to reduce the scan time by as much as 10 times by perfusing the specimen with a mixture of formalin and an MR contrast agent that reduces the spin lattice relaxation time of the tissues. Figure 1 shows a series of 1-mm-thick MR images of fixed mouse brains acquired on the 9.4T system. The details of spatial encoding in MR can be found in these good textbooks (Haacke et al., 1999; Bernstein et al., 2004). To encode the data (i.e. to image), a series of radiofrequency pulses is applied to the specimen at regular intervals (TR) ranging from 10 ms to 3 seconds. During the interval between pulses, the excited protons recover. The active stain used for specimen studies contains gadolinium (Gd), a transition metal with unpaired electrons. These paramagnetic electrons enhance the recovery process, which results in enormous gain in signal strength. The top row of Figure 1 shows images acquired with TR ranging from 20 ms to 2.56 seconds. In this formalin-fixed specimen, there is very little signal until TR reaches 320 ms. The bottom row of Figure 1 shows a specimen that was actively stained with the Gd contrast agent, and the same series of images was acquired. The faster recovery of the signal between pulses results in much stronger signal at all TRs. A plot of signal in the caudate putamen is shown in Figure 2. The time required to generate an image is TR × number of phase-encoding steps (in this example, 256 along the vertical axis). Thus, the acquisition time for images in the stained brain (lower row) is 82 seconds. In the absence of the stain (top row), the signal requires more than 2.5 seconds to recover to yield an imaging time of 656 seconds, i.e. a scan that is 8-times longer.

Figure 1.

MR images of a fixed mouse brain acquired with a repetition time (TR) of (a) 20 ms; (b) 40 ms; (c) 80 ms; (d) 160 ms; (e) 320 ms; (f) 640 ms); (g) 1280 ms; (h) 2560 ms. The top row is a specimen fixed with 10% buffered formalin. The bottom row is an actively stained specimen fixed with a mixture of formalin and ProHance. Note that the signal in the stained brain at (d) TR=160 ms is comparable to that of the unstained brain at (h) 2560 ms, which shows the enormous increase in signal realized through the active staining.

Figure 2.

Measurements of the signal in the caudate putamen in the images from Figure 1 show the exponential recovery at longer TR. Note that the tissue in the stained specimen recovers much more quickly, which allows acquisition in a much shorter time.

With a shorter TR, a specimen can be scanned with three-dimensional (3D) arrays. Again, the interested reader is directed to the literature for the technical details (Suddarth and Johnson, 1991; Johnson et al., 2007). 3D encoding strategies allow volumetric images to be acquired in which the spatial resolution is the same along any axis—the results are shown in Figure 3. The images are extracted from a 512×512×1024 3D array of an actively stained 18.5-day mouse fetus. The data are from a study of the developing fetus from embryonic day 9 through postnatal day 32 (Petiet et al., 2008). The spatial resolution is isotropic, i.e. the same along every axis. Each element of the 3D array represents the signal from a volume of tissue (voxel) that is 20 µm on a side (voxel volume=8 pl). By comparison, a clinical human brain scan can use voxels that are 1 mm on a side (voxel volume=1µl). Thus, the resolution in Figure 3 is 125,000-times greater than those typically encountered in the clinic. In the example shown in Figure 3, we have used Vitrea software that allows interactively angling the slice through the volume to define the plane in Figure 3a at ~2° from the longitudinal axis of the animal (see red line in Figure 3a). This allows viewing of the optic nerve exiting the back of the eye (arrow 1 Figure 3a). The nerve can be followed into the visual cortex (arrow 2 in Figure 3b). A set of cursors (blue and red) in this plane can then be rotated to define the oblique plane that follows the optic nerve. The result shown in Figure 3b shows the nerve entering the cranial vault (arrow 2 in Figure 3b) in the sagittal plane.

Figure 3.

Selected levels from a three-dimensional MR histology image of an E 18.5-stage mouse fetus that has been actively stained. (a) A slice interactively chosen to include the optic nerve (arrow 1). (b) A second plane perpendicular to the first has been defined (red line) to show an off-axis oblique sagittal view that allows following the optic nerve through the cranium into the brain (arrow 2). Since the resolution is isotropic, there is no loss of resolution in the off-axis planes. Spatial relationships can be easily understood using such multi-planar views, e.g. between the visual and auditory system structures such as cochlea (arrow 3). The data were acquired at 9.4T with 19.5-µm isotropic resolution.

Since the specimen is not physically sliced, i.e. the specimen is intact and the tissue has not been dehydrated, the anatomic relationships are maintained—for example, the optic nerve relative to the cochlea (arrow 3 in Figure 3a). Since the specimen is still in the cranial vault and there is no visible separation of the tissues, any linear measurements taken will be far more representative of the in vivo setting than those done with conventional optical sections.

A reasonable question to ask is why MR has been successful clinically, when computed tomography (CT) scanners provide such superb anatomical data, relatively fast, and at considerably lower capital cost than that of an MRI system. The answer lies in the superb soft tissue contrast that can be obtained in MRI. The images in Figure 4 can be described in terms a pathologist would use for tissue stains. In MRI, the signal from any tissue is dependent on the water in the tissue and how it is bound. By choosing different acquisition parameters, one can highlight different aspects of the physical interaction of the water protons with the tissue. We have coined the term “proton stain” to refer to the underlying physical mechanism that differentiates one tissue from another (Johnson et al., 1993). In this example, all attributes of the tissue water are altered to some degree by the presence of the contrast agent used as the active stain. In the T1-weighted image seen in Figure 4a, the signal is dependent primarily on differential recovery of signal between the individual radiofrequency pulses used to stimulate the tissue. In this example, TR=50 ms and TE=5.6 ms. There is also some dependence on the number of water protons in each voxel. The granular layer (GrDG, arrow in Figure 4a) is most likely dark because of a longer spin lattice relaxation time (T1). Note the strong signal from the reticular part of the substantia nigra (SNR). The boundary between this structure and the lateral terminal nucleus of the accessory optic tract (LT) in the T1-weighted image can also be resolved. The contrast in Figure 4b is dependent on the spin spin relaxation time (T2) (Sharief and Johnson, 2006). The distinction between the substantia nigra and the lateral nucleus is lost in the T2-weighted image, and the granular layer is not as evident. The substantia nigra is now dark relative to the surrounding structures. But the anterior pretectal nucleus (APT), which is not visible at all in the T1-weighted image, is now visible. In the T2*-weighted image (Figure 4c), which reflects a combined effect of T2, field inhomogeneity and susceptibility, the pretectal nucleus is no longer visible, but the medial geniculate nucleus (MG, in Figure 4c) is now seen. Vessels, not seen in either the T1- or T2-weighted images are much more prominent in the T2*-weighted image (Figure 4c).

Figure 4.

Transverse sections from 3D MR images of an actively stained mouse brain acquired with (a) T1-weighting (spin echo, TR=50 ms, TE=5 ms); (b) T2-weighting (Sharief and Johnson, 2006) (TR=200 ms, DTE=7.5 ms, 8 echoes, multiecho frequency domain image contrast [MEFIC]; (c) T2*-weighting (gradient echo, TR=50 ms, TE=5 ms, alpha=60°). Key: anterior pretectal nucleus (APT); granular layer (GrDG); lateral terminal (LT); medial geniculate (MG); substantia nigra, reticular part (SNR)

Linear morphometric measures are clearly possible for a number of structures shown in Figure 4. Since the data is isotropic, the section can be rotated interactively to yield homologous slices. Finally, since the brain is in the cranium and with no physical sectioning, the accuracy of such linear measures relative to the same measure in vivo is far greater than what can be obtained from glass slides. The data is isotropic, so volumetric measurements are possible in structures like the substantia nigra, where the boundary is well defined. This process of structural segmentation has been studied extensively in clinical radiology. Unfortunately, the boundaries of some structures are clear in one acquisition and not in another. The contrast in each image is dependent on the acquisition parameter that can be controlled, along with the physical state of the tissue (Wehrli et al., 1988; Callaghan, 1994; Haacke et al., 1999). It is worth noting that the images in Figure 4 are of the very same slice of tissue. Since the tissue is not “sectioned” physically but through the spatial encoding, it is possible to image the same tissue in many different ways. This is equivalent to having three different stains of the same section of tissue. This fourth dimension of information can help in the volumetric segmentation of the brain structure (Ali et al., 2005; Johnson et al., 2010). Figure 5 shows a 3D (volume-rendered) image in which the internal structures have been segmented using this multispectral information. Recent extensions now make it possible to segment more than 35 different structures (Badea et al., 2007; Badea et al., 2009a). Figure 5a is a volume-rendered image of a C57BL/6J mouse brain in which 37 structures have been semi-automatically segmented using a dedicated imaging pipeline. The pipeline consists of a series of sophisticated software operations that are performed sequentially through the use of scripts. In Figure 5b, the majority of the structures have been “turned off,” i.e. made invisible in the dataset to allow viewing of the relative volume and relationships between hippocampus, fimbria, fornix, amygdala, and anterior commissure. The volumes of these structures are readily computed and can be used to assess differences among strains, such as the BXD family of recombinant inbred mice (Badea et al., 2009a; Badea et al., 2009b), or to identify morphometric changes in mouse models of neurological or psychiatric conditions (Badea et al., 2010).

Figure 5.

37 structures have been automatically segmented from 3D MR images acquired at 21.5-µm resolution (Johnson et al., 2010). (a) The volume-rendered image shows many of the boundaries of these sub-volumes. (b) By rendering the overlying structures invisible, users can better appreciate the volumes, shapes and spatial relationships among deep structures such as hippocampus, fimbria, fornix, amygdala, and anterior commissure.

While the value of MR histology for neuromorphometry seems clear from the example in Figure 5, there is currently very little use of the technology in practice. This is probably due to a number of very practical barriers—no standardized protocols; a limited base of pathologists familiar with reading the material; no regulatory acceptance; the complexity of the technology; and cost. The cost for a high-field research system (7T to 9.4T) can range from $1.5 million to $3 million. These systems employ superconducting magnets that must remain cold, so there is an ongoing cost for liquid nitrogen and helium. The systems are extremely flexible, but with this flexibility comes complexity.

Figure 6 shows a possible solution recently introduced by Aspect (ASPECT Magnet Technologies Ltd. Netanya, Israel). The Aspect system employs a novel 1.0T permanent magnet that requires no cryogens and is quite easy to site. The system is 20–50% the cost of higher field systems with cryogenic magnets. The most immediate consequence of operating at lower magnetic field is a loss of signal. The signal is roughly linear with magnetic field. There are, however, some very interesting secondary effects that offset this signal loss. The T1 of tissues is considerably shorter at lower field and the T2 is longer (Johnson et al., 1985; Dockery et al., 1989; Malisch et al., 1991). These changes in the tissue parameters have significant impact on the contrast-to-noise, i.e. the distinction of one tissue from an adjacent tissue. Figure 6 shows several contiguous transverse slices from a 3D volume image of an actively stained Sprague-Dawley rat brain. While the spatial resolution (100 µm) is considerably less than that attained at higher field (43 µm in Figure 4), there is adequate resolution, signal, and contrast in this roughly one-hour scan to delineate most of the major anatomical landmarks. The protocols have been carefully designed to routinely produce scans of this quality without requiring detailed understanding of the relaxation mechanisms. These representative images allow ready definition of ventricles, cortex, corpus callosum, and several other white matter tracts.

Figure 6.

Contiguous 1-mm-thick slices from a 3D gradient echo (TR=60 ms, TE=7.6 ms, a=70°) image of an actively stained adult rat brain have been acquired at 1.0 T with 100-µm in plane resolution in 55 minutes.

Discussion

The increasing complexity of drug discovery, safety assessment, and regulatory documentation require us to consider new approaches in pathology. Current safety assessment protocols review less than1% of the available tissue. The relatively low expense of tissue preparation and the time required for an experienced pathologist to review the sections preclude any immediate change in this practice. MR histology will not replace the more traditional approaches. However, MR histology has the potential to supplement and enhance the traditional approaches. A 3D exam can be acquired on a whole, intact organ in less than one hour. The MR image provides much more accurate morphometry than can be made with traditionally prepared tissue slices, since the tissue is not dehydrated and remains intact in the cranium. Volume measurements that can be easily made from the MR images are tedious at best from physical tissue slices using a stereologic approach. Since the study is non-destructive, the same tissue can be forwarded for traditional processing (Himes et al., 2004). The MR images can provide the pathologist a survey of the entire organ by highlighting lesions that might otherwise be missed (Maronpot et al., 2004). Finally, since the 3D volumes are measured in situ with virtually no physical distortion, it becomes possible to mine much more information from tissues, such as surface areas, shapes, and structural volumes (Badea et al., 2007). For example, Cyr et al. demonstrated that changes in local morphology derived from 3D MR images reflected very subtle changes in neuronal density in animals with elevated dopamine (Cyr et al., 2005). Automated volume measurements can be used to replace tedious stereologic measures.

Many of the barriers noted to wide acceptance of MR histology still exist—limited protocols, lack of regulatory acceptance, and a limited workforce to interpret the images. But, these same barriers existed when MR imaging was first introduced to the clinical domain. Economic pressures and the need for more effective safety assessment will continue to be significant stimuli to the broad acceptance of MR histology, and it seems reasonable to expect that the technology will soon have the same broad impact it has in clinical care.

Table 1.

Magnetic resonance definitions

TR	Repetition interval between excitations of an image encoding sequence
TE	Echo time: time allowed for evolution of different spin spin relaxation
T1	Spin lattice relaxation time: characteristic time for recovery of tissue magnetization after a radio frequency (RF) excitation
T2	Spin spin relaxation time: describes the loss of signal coherence after tissue excitation. Cerebrospinal fluid (CSF) has a longer T2 than gray matter.
T2*	(Pronounced “T2 star”) Susceptibility-induced relaxation time. T2* is always shorter than T2.
isotropic	Having the same resolution along all three axes
segmentation	A process by which structures are digitally differentiated from their surrounding tissue
reconstruction	The process of generating an image from the digital signals obtained through the MR sequence
active staining	A method to reduce T1 in fixed tissue by perfusion with a mixture of fixative and appropriate MR contrast agent

Abbreviations

3D

three-dimensional

ANTS

Advanced Normalization Tools

APT

anterior pretectal nucleus

CT

computed tomography

GrDG

granular layer

ITK

Insight Segmentation and Registration Toolkit

LT

Lateral terminal

MG

medial geniculate

MRH

magnetic resonance histology

MRI

magnetic resonance imaging

MRM

magnetic resonance microscopy

NMR

nuclear magnetic resonance

SNR

substantia nigra, reticular part