| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
1 Both authors: Department of Radiology, Box 3808, Duke University Medical Center, Durham, NC 27710.
Received September 6, 2001;
accepted after revision January 24, 2002.
Presented at the annual meeting of the American Roentgen Ray Society,
Washington, DC, 2000.
Abstract
 Top Abstract IntroductionSubjects and MethodsResultsDiscussionReferences |
|---|
SUBJECTS AND METHODS. Ten patients with a diagnosis of glioblastoma multiforme underwent contrast-enhanced inversion recovery prepared three-dimensional (3D) volumetric spoiled gradient-recalled acquisition in the steady state (SPGR) MR imaging (contiguous 1.5-mm slice thickness, 96-104 slices). After this imaging sequence, each patient was brought out of the head coil into a sitting position and then repositioned in the coil. The inversion recovery prepared 3D SPGR sequence was then repeated. A commercially available software program operating on a clinical workstation was used to automatically register the second inversion recovery prepared SPGR series to the first. The speed of registration was recorded. The accuracy of each registration was estimated by recording the coordinates of eight anatomic landmarks on the registered and reference series and by calculating the mean error among matching landmarks.
RESULTS. In nine of 10 patients, the registration software produced a visually satisfactory registration. In one patient, a second registration was necessary to produce a satisfactory registration. The processing time for each iteration was 48.3 ± 3.8 sec (mean ± SD). The mean error in aligning matching anatomic landmarks ranged from 0.67 to 1.41 mm, with an overall mean of 1.18 mm. The largest error among matching landmarks was 2.3 mm.
CONCLUSION. Commercially available registration software can automatically register 3D imaging volumes in less than 1 min. The mean error in registration was approximately equivalent to the dimensions of a single voxel.
|
TopAbstractIntroductionSubjects and MethodsResultsDiscussionReferences |
|---|
We describe commercially available software that registers image volumes quickly and automatically on a clinical workstation. Additionally, in anticipation of using this registration software to monitor changes in brain tumors of patients undergoing serial MR imaging, we preliminarily evaluated the accuracy of registrations obtained with this software.
|
TopAbstractIntroductionSubjects and MethodsResultsDiscussionReferences |
|---|
MR Imaging
MR imaging of the brain was performed on a 1.5-T magnet (Signa; General
Electric Medical Systems, Milwaukee, WI) using axial T1-weighted and dual-echo
conventional spin-echo T2-weighted imaging. As part of a protocol for patients
who were undergoing intracavitary monoclonal antibody therapy, a power
injector was used to infuse 0.2 mmol/kg of gadopentetate dimeglumine contrast
agent IV at 6 mL/sec. Contrast-enhanced spin-echo T1-weighted images were
obtained in the axial and coronal planes. A three-dimensional (3D) volumetric
acquisition was then performed using T1-weighted spoiled gradient-recalled
acquisition in the steady state (SPGR) MR imaging with a preparatory inversion
pulse (TR/TE, 9.0/1.9 msec; inversion time, 300 msec; flip angle, 20°;
matrix, 256 x 192; field of view, 24 x 19 cm; number of
excitations, 1; slices, 96-104; contiguous slice thickness, 1.5 mm; imaging
time, 3.6-4.1 min). Nominal voxel size was 0.94 x 0.99 x 1.50 mm.
After completion of this pulse sequence, the patient was removed from the MR
imaging bore and brought out of the head coil into a sitting position. The
patient was then repositioned in the head coil, and the inversion recovery
prepared 3D SPGR sequence was repeated.
Image Registration
A commercially available software program (Fusion 1.0.15; General Electric
Medical Systems) operating on a clinical workstation (Ultra; Sun Microsystems,
Mountain View, CA) was used to register images from the second inversion
recovery prepared 3D SPGR sequence (referred to as the source images) to
images from the first inversion recovery prepared 3D SPGR sequence (referred
to as the reference images). Using the classification system proposed by
Maintz and Viergever [5], we
used a 3D-to-3D contour-matching algorithm registration technique that
performs global rigid-body transformations. The algorithm is automatic, with
no application of fiducial markers, manual imaging segmentation, or rejection
of suggested registrations required.
The software automatically extracts contour information from both the reference and source MR images using predefined thresholds that have been optimized for images obtained with inversion recovery prepared 3D SPGR T1-weighted sequences. Five hundred or more points along these contours chosen using a proprietary technique are used to evaluate goodness-of-fit of transformations. An optimized transformation is calculated iteratively to minimize the sum of squares of the distance between points using a classic gradient descent algorithm (Knoplioch J, personal communication).
On the basis of the optimized transformation, the source images are reformatted using trilinear interpolation to produce images (referred to as registered images) aligned with reference images (Fig. 1A,1B). This software uses an option called a "linked cursor" that allows reference imaging series to be reviewed in axial, reformatted sagittal, or coronal planes with automatic matched display of the corresponding reformatted axial, sagittal, or coronal images of the registered imaging series. After the registration is validated by manually confirming the similarity of image features at a minimum of three noncollinear points, a registered imaging series with image locations and slice thicknesses matching those of the reference imaging series is created and saved.
|
|
Evaluation of Registration Speed and Accuracy
The speed of image registration was evaluated by noting the elapsed time
between the initiation of the registration program on the workstation and the
display of matched reference and registered images using the linked cursor
option.
Two methods were used to evaluate registration accuracy. The first method was visual assessment. This method has been shown to reliably detect 2-mm misregistrations of brain MR images to CT scans [6]. Misregistrations of a similar magnitude could be detected when comparing brain MR images with MR images.
The second method used to evaluate accuracy was to measure the discrepancies in location of anatomic landmarks in the reference and registered imaging series with an image analysis tool (FuncTool 1.9p, Advantage Windows; General Electric Medical Systems). The coordinates of eight anatomic landmarks were chosen for each patient. These landmarks included the anterior and posterior commissures in all patients; the remaining six landmarks were chosen to include landmarks in the posterior fossa and on the surface of both cerebral hemispheres near the vertex. Landmarks on the periphery of the brain were favored because the misregistrations were likely to be larger at those locations than at the center [7]. Landmarks were carefully chosen to be clearly visible on both the reference and registered images and to be spatially unique (e.g., junction points of cortical veins or sulci). Because of the inherent variability of the cortical surface, these criteria did not allow the same landmarks to be chosen for each patient.
Misregistration errors for each of the eight landmarks were calculated. Accuracy of registration is expressed as the root-mean-square error, the mean error, and the maximal error for each patient.
Evaluation of Head Rotation and Translation
Because registration accuracy may be affected by the magnitude of change in
head position occurring between the times of reference and source image
acquisition, we estimated head rotation and translation using rigid-body
transformation matrices extracted from the registration software. These
parameters were calculated using two models to express changes in head
position. In the first model, changes in head position were modeled as head
rotations around the x-, y-, and z-axes followed by
translations were calculated using previously described algebraic equations
[8].
The order of rotations and translations must be exactly specified using the first model. The magnitude of rotations and transformations derived from this model differ if the order of rotations and translations is altered. Because of this drawback, we used a second model to derive a single rotation and translation to describe changes in head position. In this model, changes in head position were modeled as a single simultaneous rotation around and translation along a fixed axis. The fixed axis in this model is generally oblique to all the cardinal axes of the coordinate system. For this model, matrix logarithms were used to decompose the rigid-body transformation matrix into a set of simultaneous rather than sequential rotations and translations. The single angle of rotation around the fixed axis was derived from the length of a vector composed of the instantaneous elementary rotations around the x-, y-, and z-axes, as previously described [8]. The single translation was derived from the length of a vector composed of the instantaneous elementary translations in the x, y, and z directions. This translation represents the distance traveled by the origin of the source coordinate system during the simultaneous rotations and translations [8].
The accuracy of the rotations and translations reported for both methods depends on the accuracy of the registration matrix calculated by the registration software. Because the registration matrix is not perfectly accurate, the reported rotations and translations should be regarded as estimates only.
|
TopAbstractIntroductionSubjects and MethodsResultsDiscussionReferences |
|---|
In nine of 10 patients, the image registration software produced a visually accurate registration (Fig. 2A,2B,2C). For one patient, the software did not produce a visually accurate registration. In this case, the registered images were used as source images for a second registration, which produced a visually accurate registration (Fig. 3A,3B,3C,3D). The set of registered images produced in the second registration step was used in the assessment of registration accuracy, and a net transformation matrix equal to the dot product of the transformation matrices in the two registration steps was used in the assessment of head motion and rotation.
|
|
|
|
|
|
|
Analysis of discrepancies in the coordinates of anatomic landmarks showed that the errors ranged from 0.1 to 2.3 mm. The mean error for each patient varied from 0.67 to 1.41 mm. The overall mean error for the entire patient group was 1.18 mm, with a standard error of the mean of 0.04 mm. The root-mean-square error for each patient varied from 0.68 to 1.49 mm. The mean root-mean-square error for all patients was 1.26 mm.
Using the first model to describe changes in head position, we estimated that the magnitudes of rotations around the x-axis (also called pitch) ranged from 0.1° to 7.8°, with a mean rotation magnitude (±SD) of 2.3° ± 2.4°. The magnitudes of rotations around the y-axis (also called roll) ranged from 0.1° to 2.7°, with a mean rotation magnitude of 1.5° ± 0.9°. The magnitudes of rotations around the z-axis (also called yaw) ranged from 0.2° to 20.8°, with a mean rotation magnitude of 5.0° ± 7.2°. The magnitude of translations along the x-, y-, and z-axes was 2.7 ± 3.0 mm, 0.7 ± 0.6 mm, and 3.9 ± 4.6 mm, respectively, with a maximal observed value of 14.6 mm. The patient whose initial registration was visually unsatisfactory (Fig. 3A,3B,3C,3D) had the largest rotation in head position along any axis and the second largest translation (10.0 mm).
Using the second model to describe changes in head position, we estimated that the magnitude of rotations around a single axis ranged from 0.3° to 21.7° with a mean (±SD) of 6.3° ± 7.0°. The translation of the origin along the single axis ranged from 0.6 to 14.8 mm with a mean of 5.6 ± 4.6 mm.
|
TopAbstractIntroductionSubjects and MethodsResultsDiscussionReferences |
|---|
Traditionally, brain tumors are monitored by MR imaging, using sequences
with relatively large (5-7 mm) slice thicknesses and small interslice gaps.
Three-dimensional volumetric MR imaging can be advantageous because image
slice thickness can be smaller (
1.5 mm) and no gaps are present. Two
problems make comparison of 3D volumetric MR images obtained at different
times difficult. First, because the positioning of the patient's head is not
completely reproducible, the orientation of image slices in space and the
position of the slices are generally not consistent over time. Second, the
amount of image data in large multislice volumetric series may make comparison
of serial images unwieldy. We chose to investigate image registration in
patients with brain tumors monitored using 3D MR volumetric imaging because
registration software offers a possible solution to both these problems.
Our study found that this commercially available software program produced automatic registrations on 3D volumetric MR imaging series containing between 96 and 104 images in less than 1 min. No radiologist input was needed for image segmentation.
We evaluated the accuracy of the registration software using two techniques. First, visual evaluation was used to provide a preliminary assessment of the registration. Although this type of evaluation may seem suboptimal because it is qualitative and subjective [7], prior studies have shown that misregistrations as small as 2 mm can be reliably detected [6, 10]. We found visually satisfactory registrations in nine of 10 patients. In another patient, re-registration of the image data yielded a visually satisfactory result. This case illustrates that registration software may perform properly for most patients but may on occasion produce poor registration [11], particularly when source images are poorly aligned with reference images. Visual evaluation continues to play an important role in quality assurance by detecting registration failures, particularly when registrations are automatic.
Second, we quantitatively evaluated the accuracy of registration by measuring the discrepancies in location of anatomic landmarks in the reference and registered images. As noted by Fitzpatrick [11], registration accuracy should be evaluated in the context of the clinical task to which it is being applied. For example, the accuracy of a given registration technique applied to measuring changes in the size of the ventricular system may differ from that of the same technique applied to measuring changes in the size of convexity meningiomas. Although many measures of registration accuracy have been proposed, Fitzpatrick recommended the target registration error (the disparity in the positions of two corresponding clinically relevant points after registration) as the preferred measurement.
In our study, eight anatomic features in each patient were chosen to simulate targets of brain tumor diagnosis. Features in the posterior fossa and along the convexities of the brain bilaterally were included in each patient for two reasons: these locations represented realistic possible targets for our application; and misregistrations, because of angular malalignment, are likely to be larger at these locations than at the center [7]. By choosing points at a distance from the center, we attempted to avoid underestimation of registration error. Anatomic features whose locations were clearly ascertainable on both reference and registered images were chosen because any uncertainty in identifying the locations of the features contributes to overestimation of registration error.
Using this technique, we found that the mean error for each patient varied from 0.67 to 1.41 mm, and the root-mean-square error varied from 0.68 to 1.49 mm. The overall mean error for the entire patient group was 1.18 mm, which is roughly equivalent to the average dimension of a single voxel. Some previously reported registration algorithms have comparable accuracy [12], whereas some are notably more accurate. For example, some algorithms produce registrations with subvoxel accuracy [13], although these algorithms generally require segmentation of the brain from extracerebral structures and may be less rapid. Although the level of accuracy of the algorithm under study is not adequate for detection of submillemeter changes in the size of brain structures, this level of accuracy may be adequate for the evaluation of changes in the size of brain tumors.
Two limitations to our estimate of registration accuracy are apparent. First, because the reference and source images were obtained only minutes apart and on the same scanner, our results may not be applicable to the task of registering typical clinical images, which are acquired at longer intervals (during which machine drift can occur) or on different scanners. Second, because positions of anatomic landmarks were assessed on planar images, the co-ordinate values in the z direction were not continuous. The mean error in the z direction was slightly smaller than the mean errors in the x and y directions (0.42 mm vs 0.51 and 0.65 mm, respectively), suggesting that this limitation may have resulted in overestimation of registration accuracy.
Because variation in head position between the times of acquisition of the reference and source imaging series may affect the quality of registration, we estimated this variation using two models in our patient group. The maximal observed value of head rotation angle in our patient group during the same imaging examination was larger than that seen on prior studies of patients with brain tumors monitored by serial MR imaging examinations. Nelson et al. [12] reported maximal rotations of less than 10° around any principal axis on the images of 12 patients with brain tumors. We found maximal rotations of 13.7° around any principal axis on images of 30 patients with intracerebral glioma (Barboriak et al., presented at the American Society of Neuroradiology, April 2001). It is somewhat encouraging that in our current study the initial attempt at registration failed in only one patient whose head position varied more than that typically seen in these patient groups.
Several technical factors require careful evaluation in future studies before the combination of image registration and thin-section volumetric gradient-echo MR imaging can be validated for use in monitoring brain tumors. Factors related to the image registration procedure may lead to decreased sensitivity to change in brain tumors. The contour-matching algorithm used by this registration program performed reasonably well when registering volumetric imaging series obtained in a single imaging session. However, it remains unclear whether this level of performance would be achieved when registering series obtained at different imaging sessions. Interval craniotomy, hemorrhage, or changes in mass effect may result in poorer quality registration because of alterations in image contours, leading to decreased sensitivity to small changes in tumor size. In addition, visualization of small lesions may also be limited because of the image interpolation required to reformat source images into registered images. The software evaluated in this study used trilinear interpolation, a method that is computationally efficient but is less accurate than other forms of interpolation [14].
Factors related to the inversion recovery prepared 3D SPGR acquisition may also limit sensitivity of volumetric imaging to small changes in brain tumor enhancement. Clinical studies have suggested that the signal-to-noise ratio and the conspicuity of enhancing lesions is decreased on SPGR images compared with conventional spin-echo T1-weighted images [15, 16]. A more recent analysis showed that signal-to-noise ratio on a thin-slice T1-weighted volumetric gradient-echo technique was essentially equivalent to that of spin-echo images (allowing for the reduced slice thickness in the volumetric technique) and that small lesion visualization was superior when the volumetric thin-slice technique was used [17]. Our thin-slice technique performed after IV infusion of 0.2 mmol of gadolinium should be relatively sensitive to small changes in glioma enhancement. However, this assumption will need to be tested by directly comparing the sensitivity of inversion recovery prepared SPGR imaging with that of conventional T1-weighted spin-echo imaging.
In summary, this study describes commercially available software that can produce rapid, automatic, and reasonably accurate registrations of volumetric brain MR imaging series on clinical workstations. Such developments represent an early step toward the integration of registration algorithms into clinical imaging.
Acknowledgments
We thank Gopal Sundaramoorthy, Jerome Knoplioch, and Jean Lebarre for their
assistance with the Fusion software.
|
TopAbstractIntroductionSubjects and MethodsResultsDiscussionReferences |
|---|
CiteULike 
Complore 
Connotea 
Del.icio.us 
Digg 
Reddit 
Technorati What's this?
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
