Atrophy

3.6.1

Historical Perspective

Since the early days of X-ray computerized tomography (CT) in the late 1970s, it has been possible to visualize the spinal cord in cross-section in vivo, and to make quantitative measurements of the cord diameter. Early studies required the introduction of an iodinated contrast agent into the cerebrospinal fluid (CSF) by means of a lumbar puncture, in a procedure known as computed myelography. Cord atrophy was clearly visible, and the normal range of cord size in the cervical region as measured on CT was established early on ( Figure 3.6.1 ).

Early quantitative measurements made from magnetic resonance imaging (MRI) scans also established the normal range of cord diameters in the cervical and thoracic regions. MRI has considerable advantages over CT scanning: there is no ionizing radiation involved, and the natural contrast between the cord and the CSF-filled space is much greater so that contrast agent in the arachnoid space is not needed. Whether from CT scans or from MRI, linear size measurements in the anterior-posterior (A-P) and left-right (L-R) directions were made manually by either using a rule held against a radiographic film or using a computer image analysis system. Computerized systems were also used to draw around the cord in order to assess its cross-sectional area.

While the measurement of cross-sectional areas was shown to be feasible, manual delineation of the cord raises the issue of whether such measurements are reliable. It was recognized early on that human perception of the location of the boundary between the cord parenchyma and the CSF is heavily influenced by the brightness and contrast settings of the radiographic film or digital display, and that standardization was needed if assessment consistency was to be achieved. The development of standardized methods for quantitative analysis, and preferably at least some automation, aims to reduce the variability of repeated measures by a single assessor (intraobserver reproducibility) or between different assessors (interobserver reproducibility). If a method can be developed that is completely automated, then the only influence on the measurements is the repeatability of the scanning technique (the so-called scan-rescan variability). An evaluation of all of these types of error is needed for a spinal cord atrophy measurement technique to become established and accepted as reliable.

This also brings up the question of how reliable such measures need to be in practice, which is of course dependent on the context of the measurement. Changes in spinal cord cross-sectional area may occur at a very slow rate, of the order of a few percent per year, which makes challenging the detection of change in relatively short longitudinal studies, or in cross-sectional studies at the early stages of disease. In order to keep patient sample sizes manageably small, it is imperative to employ the most reliable and robust techniques that are currently available, and to continue to improve on existing methods.

3.6.2

MRI Scanning Considerations

Successful segmentation of the spinal cord relies on there being a good demarcation of the boundary between the cord parenchyma and the surrounding CSF. In T ₁ -weighted imaging, the CSF will be darker than the cord due its long T ₁ , while in T ₂ -weighted imaging, the CSF will be brighter due to its long T ₂ . Either type of scan can therefore, in principle, be employed in atrophy measurements. In order to recommend one particular type of scan, other considerations must therefore be brought into play.

First is the spatial resolution that can be achieved. The highest spatial resolution is normally achieved by using a three-dimensional (3D) acquisition, with two spatial dimensions being phase encoded and one dimension being frequency encoded. This allows the MR signal to be acquired from the whole imaging volume throughout the imaging sequence, maximizing signal-to-noise ratio (SNR) and leading to the best possible spatial resolution. This type of acquisition requires a very short repetition time (TR) that is really only compatible with gradient-echo pulse sequences and T ₁ -weighted imaging. While it is possible to perform 3D acquisition of T ₂ -weighted images, 3D T ₂ pulse sequences are still normally less SNR efficient than T ₁ -weighted sequences, and they rely on multiple-echo rapid acquisition with relaxation enhancement (RARE) with very long echo trains, which are prone to blurring fine detail, and artifacts at feature boundaries.

However, T ₂ -weighted imaging may be desired for other reasons, for example to look for focal hyperintensities in the cord that are indicative of pathology. One potential problem is that pathology causing hyperintensity near the edge of the cord could confound segmentation algorithms, leading to an underestimation of cord area in that region. Because of the shape of the spinal cord, with only slow changes in the cross-section along its length, it is feasible to assess atrophy on T ₂ -weighted multislice cord images acquired axially and with relatively poor resolution in the slice (head-foot) direction. Nevertheless, care must be taken in regions where the cord does not pass perpendicularly through the image plane, because where the cord passes at an angle through the slice, the boundary will appear blurred because the boundary position varies through the image slice.

3D T ₁ -weighted acquisition is normally performed with the frequency-encoded direction being head-foot, and the stored image planes are therefore either sagittal or coronal. Since the cord cross-section is viewed in the axial plane, the scans must therefore be reformatted into axial, and at this point it is prudent to use a multiplanar reconstruction (MPR), with free selection of the resampled plane orientation, to make sure that the cord axis is truly perpendicular to the reconstructed planes at the most important anatomical location, such as the cervical region. Loss of image quality on reformatting of 3D acquired images can be completely avoided if sinc interpolation is used.

MR images are prone to motion artifacts, and the cervical region of the cord can suffer particularly badly. Because of the way that MR image data are collected, gross motion of the patient often appears as replications of the image features, or “ghosts”, that are superimposed on top of the image at intervals over the field of view. The main causes are breathing, swallowing, or movement caused by discomfort of the patient during long scans. In particular, in the thoracic region, cardiac motion and pulsation of the major vessels are issues. The artifacts caused by breathing, swallowing, and cardiac motion can be greatly reduced by using “saturation bands” that cover the anterior portion of the chest and abdomen away from the cord. In saturation bands, radiofrequency (RF) pulses are repeatedly applied in conjunction with a magnetic field gradient to select and dephase the signal from the moving tissue so that it cannot corrupt the region of interest. CSF pulsation can cause bright signal that is misplaced and can appear

superimposed on the cord. Flow-compensated sequences greatly alleviate the problem, and these work by “gradient moment nulling” so that spins moving with a fairly constant velocity do not experience a phase shift relative to stationary spins, and therefore do not appear spatially misplaced.

3.6.3

Semiautomated Atrophy Measurement

All methods for estimating spinal cord atrophy from MRI scans have, at their heart, a way of determining the location of the boundary between the cord parenchyma and the surrounding material. For much of the length of the cord, it is bordered by CSF, although there are many anatomical levels where the cord surface touches the surface of the spinal canal. This leads to a loss of the distinctive boundary as the surrounding tissues may have a very similar intensity to the cord. Furthermore, although the cord is roughly ovoid in cross-section, the shape changes from a more circular shape in the upper cervical region to a flattened elliptical shape in the mid-lower cervical region, becoming more circular again in the thoracic region before tapering out to the conus in the upper lumbar region ( Figure 3.6.2 ). With the high spatial resolution that is now available from modern MRI scanners, the dorsal and ventral nerve roots that come from the cord at regular intervals, passing through the intervertebral foramina, are increasingly visible on thin-slice images and can further confound segmentation software. Nevertheless, the cord is a long, thin, curved cylinder of irregular cross-section, with a convex or only slightly concave boundary and with a cross-sectional shape that changes only slowly along its length. These properties can be used to good effect when segmenting the cord, and can help to overcome the problems associated with marginal spatial resolution and poor and inconsistent contrast.

Cord segmentation methods can be divided roughly into those that use cord shape information and those that do not. The methods that do not use shape information have tended to concentrate on measuring the cord at a specific anatomical level, particularly in the midcervical region. This is because, with normal neck flexion, at this level the spinal cord is usually surrounded completely by CSF, or there is only a small region where the cord abuts the border of the spinal canal. There is good cord–CSF contrast, and the division between them involves the selection of an intensity threshold.

3.6.3.1

Intensity Threshold Selection

With the spatial resolution of the order of 0.8–1.0 mm that is achievable on modern MRI scanners, the shape of the cord is easily judged by the human eye. Nevertheless, at the boundary between the cord and the CSF, there is a so-called partial-volume effect that blurs the boundary because an image voxel contains a mixture of parenchyma and CSF, with the proportion of each being determined by the position of the boundary within the voxel ( Figure 3.6.3 ).

When an image pixel contains a 50% mixture of cord and CSF, the intensity should be midway between the intensity of pixels that contain pure cord and that for pure CSF. Thus, the threshold to separate the two is logically set halfway between these intensities. This forms the basis of the cord area measurement method developed by Losseff et al., which involves selection of the intensity threshold by defining two regions of interest (ROIs), one surrounding just the cord and the other surrounding the whole of the intravertebral space. With the intensity threshold set, it should be a simple matter to outline the cord by following a contour of isointensity at the cord surface. Contour following may be problematic when the cord touches the surface of the spinal canal, with the contour “spilling out” into adjacent structures; hence, the Losseff method is specifically used at the C2–C3 level, where the cord is usually centered within the canal.

3.6.3.2

Edge Detection

The location of the edge of the cord may also be determined in another way. The intensity difference between the cord and CSF, together with the inherent blurring and finite pixel size in MR images, means that there is a gradient in intensity at the boundary. The true edge position is therefore most likely to be the point of maximum intensity gradient magnitude. This definition of the edge has the advantage that it is relatively insensitive to small variations in image intensity over the image field of view, such as occur due to the nonuniform reception properties of the RF coils. The spinal cord can be segmented over an extended region at the C2 level using edge detection, as was demonstrated by Tench et al. ( Figure 3.6.4 ).