British Journal of Radiology 75 (2002),S31-S35 © 2002 The British Institute of Radiology
The application of PET–MR image registration in the brain
R Myers
Imaging Research Solutions Limited (IRSL), Cyclotron Building, Hammersmith Hospital, DuCane Road, London W12 0NN, UK
 |
Abstract
|
|---|
The coregistration in three-dimensional space of positron emission tomography (PET) and magnetic resonance (MR) image volumes has, over the last decade, become a matter of routine in the analysis of brain PET studies. The ability to objectively localize small regions of interest in PET using images more closely correlated to tissue structure has itself improved the effective resolution of PET. There are a number of highly effective software packages for image coregistration available in the public domain. Voxel-by-voxel coregistration, involving little or no intervention from the user can, on today's computing hardware, provide fast and accurate registration with little or no pre-processing and algorithms based on mutual information measures now seem to be the mathematical method of choice. Registration may be applied in a number of ways. Rigid body registration is used to match a single subject's brain scanned using either different imaging modalities or as serial scans with the same modality. Increasingly, this technique is being extended to studies of disease involving regional atrophy, where location and extent of tissue loss can be identified. Non-linear registration can be used to warp a subject's brain onto a template, atlas or other standardized guide. While numerous examples are available of the added value produced by image registration in the brain, similar examples are not yet available from registration in the torso, where the problem is much more complex. It is here that newly emerging hardware such as combined PET/CT scanners may prove their worth.
|
Introduction
|
|---|
The increasing use of multi-modality medical imaging has generated a concomitant increase in the use of image registration techniques to fuse pairs of images. PET–MR registration in the brain represents only one small part of the field of image registration. However, some of the issues involved may be more generally applicable to other modalities such as CT, ultrasound and SPECT and to registration in the body. An excellent review of the wider field of image registration can be found in Maintz and Viergever [1].
It is important to recognize that, while PET images may appear to contain representations of anatomical structures, a PET image maps a distribution of some function in the brain, rather than anatomy itself. It may therefore be necessary to co-register the PET functional image with an anatomical image for accurate placement of regions of interest, for correlating functional and structural information or for correction of partial volume loss. Although these three applications are the commonest reasons for co-registering brain images, the technique may be required for other purposes such as targeting therapy or image-guided surgery, particularly in oncology whole body imaging. It is easier to register images of the brain than of the body as the shape and size of the brain does not normally change over short periods of time, allowing it to be treated as a "rigid-body". Thus, in the absence of pathology, a given subject coming back for a second scan can be considered to have a brain that will be the same size and shape as in the first scan.
There are a number of techniques used for co-registration. They fall into the two broad categories of primary and secondary registration. Primary registration refers to ways of acquiring the data such that they are in the same space as an earlier scan, which might be regarded as "preregistration". Examples would be the use of stereotaxic frames to position the patient in a reproducible way, and scanner gantry repositioning. Secondary registration is distinct from this in that the data are reprocessed after acquisition of the scan. A pair of scans can be registered either by using markers placed in the field of view or by matching internal landmarks or surfaces already present in the image. However, the technique of choice for registration of brain images utilizes the matching of overlapping voxels in the images and a number of mathematical methods have been employed to this end.
The process of registering two or more images can be divided into five steps. The first is pre-processing, which ideally should be kept to a minimum. The second is the registration itself, which produces a calculation of the transformation parameters relating the two images. Then follows transformation, which uses the output of the registration to rotate, translate and re-slice one image volume into the space of the other. The last two steps, the importance of which are paramount and yet are often missed, are the assessment and the display of the registered image. Each of these steps will now be discussed in more detail.
|
Pre-processing
|
|---|
PET images often have a very low information content. An example is shown of a normal scan acquired after administration of carbon-11 labelled raclopride, which binds to D2 receptors in the brain with high uptake in the striatum (Figure 1
). The image shown in Figure 1a represents the final endpoint of the study, a parametric map of binding potential (BP) of [11C]raclopride in the brain. However, this image volume, as it stands, does not contain sufficient information to guarantee a good registration to MR. A solution to the problem is to add together all frames from the entire dynamic scan, to produce a summed image volume (Figure 1b). This contains sufficient information to achieve satisfactory co-registration, after which the transformation parameters acquired from this coregistration are applied to the BP map and the summed images are discarded.
View larger version (59K):
[in this window]
[in a new window]
|
Figure 1. Pre-processing—improving information content. Image (a) shows a binding potential map of [11C]raclopride in the brain of a normal subject. While the desired endpoint of the coregistration would most likely be the ability to fuse this with its corresponding MRI, the information content of the image is insufficient to give a good result. Image (b) shows a summed image of the whole dynamic scan used to calculate image (a). This now contains enough information to achieve satisfactory calculation of the transformation parameters which can then be applied to the binding potential map.
|
|
|
Registration
|
|---|
The most commonly used paradigm for registration is voxel matching, where some measure of voxel intensity similarity is optimized to achieve automated registration. The underlying principle behind this technique is illustrated in Figure 2. A scattergram of the voxel values in the first image against the voxel values in the second image when the images are identical and perfectly registered yields a straight line (Figure 2a). In Figure 2b, one of the images is moved out of register by a single voxel and the relationship between the voxels in the two images is no longer linear but forms a distinctive scatter pattern. Voxel-based registration works to minimize this scatter and there are a number of different algorithms using this principle available in the public domain.
View larger version (60K):
[in this window]
[in a new window]
|
Figure 2. Voxel based coregistration. If the voxel values from exactly registered pairs of identical images are plotted against each other, a straight line is produced, as shown in (a). In (b), one image has been translated in the Y direction by a single voxel and the scattergram now has a characteristic pattern. Voxel based coregistration algorithms work to minimize the spread of this distribution. The images shown here were produced using ANALYZE AVW version 3.1 [5].
|
|
The first algorithm that successfully used voxel matching for registration was developed by Woods and co-workers in the early 1990s. The algorithm used calculated the ratio of one image to the other on a voxel-by-voxel basis and then iteratively moved the images relative to one another to minimize the variance of this ratio. This technique proved remarkably accurate both within modality [2] and across modalities [3], although some difficulties were encountered in situations where a structure giving a high signal in one image had no correlate in the other. This was most notable in the case of the high scalp signal in MRI images for example, for which there is no match or correlate in a PET blood flow image. An additional pre-processing step was required to remove the scalp from the MR.
An alternative and very robust approach was developed by Studholme and others in Dave Hawkes's Image Processing Group at Guys and St Thomas's Hospitals. Known as MPR and derived from communication theory, it maximizes the mutual information in the two images, thus minimizing the information, or joint entropy, in the combined image [4]. This algorithm is both accurate and less susceptible to the problems encountered with non-correlating structures.
|
Transformation
|
|---|
The output from the registration process comprises a set of translation, rotation and scaling parameters, which, in combination, allow one image to be spatially transformed into the space of the other. If the objects in the two images have the same shape, for example the brain of the same individual scanned with different modalities, this is known as "rigid body transformation". At the opposite extreme, one image may be warped into a completely different shape such as occurs when the image of a subject is transformed onto a template derived from one of the various stereotactic atlases available for statistical mapping. This is called "elastic transformation". Between these two extremes are a range of other transformation types allowing various degrees of scaling.
There are some issues to be considered before undertaking the image registration process. Firstly there is the discrepancy in image size between PET and MR. Most PET images are 128 x 128 voxels in the X and Y dimensions while MR images are usually 256 x 256. Choosing to register the smaller PET image to the larger MR image will retain the high resolution in the MR image but at the expense of creating a lot of interpolated data in the registered PET image. The converse maintains the quantitative accuracy of the PET data at the expense of the MR resolution. Which is the correct approach will depend upon the application and can be best judged by the user.
Accurate measurement of the voxel sizes, the distance in space represented by each voxel in the original image data, is essential. The data can only be interpolated correctly if the voxel measurements of the set of images to be registered are known. These should be measured in the scanner under the same set of conditions used for the image acquisition and the values given by the manufacturer should not be relied upon. If accurate measurements are not available, it is possible to use a registration algorithm that allows for scaling in the X, Y and Z directions as well as translations and rotations, a so-called 9-parameter fit.
A further consideration is the interpolation of the resliced image. Interpolation occurs when one image has to be scaled and/or rotated relative to another and, as a result, voxels are added or removed. Linear interpolation is a commonly used method as it is computationally simple. The process is illustrated in Figure 3. If the two rectangles shown are rotated, requiring additional voxels to be inserted at their junction, then linear interpolation takes the mean of adjacent voxels. Averaging the values in the adjacent orange and grey voxels created the green voxels in Figure 3b. The resulting image has smooth boundaries with reduced noise, which gives a pleasing appearance. However, if it is important to maintain the signal to noise ratio or the original quantitative values of all voxels after reslicing, then nearest neighbour interpolation is more appropriate. In this case, the additional voxel values are created by replication of adjacent values, as in Figure 3c. This gives rise to a less pleasing, more "pixelated" image but is more quantitatively accurate. These are the two methods of interpolation most often used in PET. However, in MRI the interpolation of choice is usually interpolation using a sinc function which is more computationally expensive but provides a better representation of the image characteristics.
View larger version (31K):
[in this window]
[in a new window]
|
Figure 3. Interpolation occurs when voxels are inserted as a result of rotation or scaling of an image. Images (b) and (c) show image (a) rotated through 45 degrees, using linear and nearest neighbour interpolation, respectively. Linear interpolation inserts voxels along the junction of the orange and purple areas which have a value which is the mean of those in the outer areas, as can be seen on the colour bar. Nearest neighbour interpolation uses the values already existing in the outer areas, which has the effect of giving a pixelated, less smooth appearance.
|
|
|
Assessment of the registered images
|
|---|
Voxel-based coregistration algorithms are designed to manipulate the input images in order to satisfy some predetermined mathematical criterion and different algorithms use different optimization schemes to achieve this. The robustness of an algorithm reflects how frequently the optimized result generates perfect registration, and sadly no algorithm can guarantee this 100% of the time. In addition, incorrect input to the program, such as attempting to coregister a sagittal volume with a transverse volume, will generate a result which will be nonsense! It is thus possible to achieve an apparently error-free registration which does not result in well aligned images and there is hence a critical need to assess the results visually. Key to this assessment is the recognition that the problem is three-dimensional and visualization must be carried out in all three orthogonal directions.
|
Display
|
|---|
A number of tools exist for displaying coregistered images, both for assessment and final display. For the purposes of assessment of goodness of coregistration, a most effective method is to use autotracing of contours on registered pairs of images displayed side by side. Alternatively, the images may be overlaid while adjusting the proportions of each image in the fused result. Crosshairs can also be used in some packages to assess how closely the images match at any given point. Either method requires care on the part of the operator as areas of hyperintensity and hypo-intensity may not correlate precisely, for instance ventricles and white matter may appear of equal intensity on a PET scan which they obviously do not on MRI. Some software packages allow the images to be moved around to assess whether the coregistration can be improved if re-applied.
In summary, PET/MR registration in the brain can operate routinely to a high degree of accuracy. Robust algorithms are available which allow registered images to be routinely generated but they rely on the input of good quality data, which may require pre-processing. The assessment of the final result in order to check the accuracy of the registration is the responsibility of the operator, who is potentially the weakest link in the registration process. Registration is as important in the body as in the brain but is more difficult because the body cannot be regarded as a rigid body. Other solutions may be needed for registration outside the brain but some of the lessons learnt from brain registration may inform this process.
|
References
|
|---|
- Maintz JBA, Viergever MA. A survey of medical image registration. Med Image Analysis 1998;2:1–37.
- Woods RP, Cherry SR, Mazziotta JC. Rapid automated algorithm for aligning and reslicing PET images. J Comp Assist Tomogr 1992;16:620–33.[Medline]
- Woods RP, Mazziotta JC, Cherry SR. MRI-PET registration with automated algorithm. J Comp Assist Tomogr 1993;17:536–46.[Medline]
- Studholme C, Hill DLG, Hawkes DJ. Automated three-dimensional registration of magnetic resonance and positron emission tomography brain images by multiresolution optimization of voxel similarity measures. Med Phys 1997;24:25–35.[Medline]
- Robb RA, Hanson DP, Karwoski RA, Larson AG, Workman EL, Stacy MC. ANALYZE: A comprehensive, operator-interactive software package for multidimensional medical image display and analysis. Comp Med Imaging Graph 1989;13:433–54.
This article has been cited by other articles:
|
|
|
|
 
H.-Y. Lee, Z. Li, K. Chen, A. R. Hsu, C. Xu, J. Xie, S. Sun, and X. Chen
PET/MRI Dual-Modality Tumor Imaging Using Arginine-Glycine-Aspartic (RGD)-Conjugated Radiolabeled Iron Oxide Nanoparticles
J. Nucl. Med.,
August 1, 2008;
49(8):
1371 - 1379.
[Abstract]
[Full Text]
[PDF]
|
|
|
|
|
|
|
J. Andrews and J. C. Mason
Takayasu's arteritis--recent advances in imaging offer promise
Rheumatology,
January 1, 2007;
46(1):
6 - 15.
[Abstract]
[Full Text]
[PDF]
|
|
|
|
|
|
|
V. Krishnasetty, A. J. Fischman, E. L. Halpern, and S. L. Aquino
Comparison of Alignment of Computer-registered Data Sets: Combined PET/CT versus Independent PET and CT of the Thorax
Radiology,
November 1, 2005;
237(2):
635 - 639.
[Abstract]
[Full Text]
[PDF]
|
|
|
Top
Abstract
Introduction
