| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
1
Diagnostic Radiology Department, The Clinical Center, National Institutes of
Health, Bldg. 10, Rm. 1C660, Bethesda, MD 20892-1182.
2
Department of Radiology, Uniformed Services University of the Health Sciences,
4301 Jones Bridge Rd., Bethesda, MD 20814-4799.
Received March 31, 2000;
accepted after revision August 2, 2000.
Address correspondence to P.L. Choyke.
Abstract
 Top Abstract IntroductionSubjects and MethodsResultsDiscussionReferences |
|---|
SUBJECTS AND METHODS. Twenty-one patients scheduled for isolated liver perfusion therapy for metastatic disease underwent contrast-enhanced three-dimensional MRA to determine vascular anatomy. Vascular anatomy was validated at the time of surgery. We displayed the image data, using five techniques: maximum intensity projection, targeted maximum intensity projection, isointensity surface (isosurface), connected isointensity surface (connected isosurface), and ordered region growing skeleton (skeleton). Four observers, blinded to the surgical results, interpreted each technique in random order without patient identifiers. Areas under the ROC curves, kappa values of interobserver variability, and time to interpret each display were compared.
RESULTS. Skeletonized MRA had the highest area under the ROC curve (Az, 0.90 ± 0.04) compared with the other techniques (p < 0.013). Kappa scores of agreement were also highest for skeletonized MRA (0.75 ± 0.04) and had no overlap at the 95% confidence level compared with other techniques. Compared with source images, all visualization methods were faster to interpret, but the skeleton technique was more quickly (p = 0.04) interpreted than the other techniques.
CONCLUSION. Skeletonized MRA with the skeleton connectivity algorithm is a semi-automated method of displaying complex arterial anatomy. Compared with other techniques, it is more accurate, more consistent among observers, and slightly faster to interpret. Skeletonization should be applicable to CT angiography and MRA.
|
TopAbstractIntroductionSubjects and MethodsResultsDiscussionReferences |
|---|
There are several alternatives to the maximum intensity projection. One is the isosurface display that generates a three-dimensional (3D) surface display on the basis of a predetermined threshold value [3]. This technique can be improved by selecting an origin (seed point) and requiring that any generated surface must be in continuity with the seed point (connected isosurface), but otherwise the technique is identical to the isosurface display [3, 5]. Another vessel-mapping technique based on the skeleton technique has recently been proposed [6]. In this method, the user defines both the start and endpoint of the vessel, and an algorithm generates the most likely pathways connecting them. A skeleton of the vascular tree depicting the vessel's centerline is produced. The skeletonized angiogram can be superimposed on the maximum intensity projection or viewed separately as a 3D shaded-surface display. Skeletonization highlights vessels that might otherwise be inconspicuous on the maximum intensity projection. Individual vessels can be selectively color encoded to simplify display.
In our study, we compared five methods of depicting hepatic MRA data for accuracy, inter-observer agreement, and speed of interpretation.
|
TopAbstractIntroductionSubjects and MethodsResultsDiscussionReferences |
|---|
The MRA protocol was as follows: all scans were obtained on a 1.5-T MR unit with a peak gradient strength of 23 mT/m and a maximal slew rate of 120 T·m-1·sec-1 (Echospeed; General Electric Medical Systems, Milwaukee, WI). A 20- to 22-gauge IV catheter was placed in an arm vein before the patient entered the scanner. The catheter was attached to a dual-headed power injector (Spectris; Medrad, Pittsburgh, PA) loaded with a nonionic MR contrast agent, gadoteridol (ProHance; Bracco Diagnostics, Princeton, NJ). The injector was set to inject contrast agent at a rate of 2 mL/sec for a total of 0.2 mmol/kg (22-40 mL) followed by a 20-mL flush of 0.9% saline at 2 mL/sec. This injection rate was selected because it approximately matched the time of acquisition yet did not cause susceptibility or K-space modulation effects [1]. Breath-hold 3D spoiled gradient-echo imaging (TR/TE, 5.2/1.4; flip angle, 45°; field of view, 35-42 cm; bandwidth, 32 kHz; acquisition, 1; matrix size, 256 x 128; slices, 28 in 34 sec) was performed. Lines of K-space were acquired sequentially so that the center of the K-space was in the middle of the acquisition. Each slice was 2.5- to 4-mm thick depending on the size of the patient. In all patients, acquisition began after a standard delay of 15 sec from the beginning of the injection.
Image Analysis
Each entire 3D data set was composed of a series of 56 coronal source
images. The following image analysis algorithms were applied to the source
images (Figs.
1A,1B,1C,1D,1E,1F
and
2A,2B,2C,2D,2E,2F):
maximum intensity projection, targeted maximum intensity projection,
isosurface, isointensity surfaces constrained with a starting seed placed over
the origins of the celiac and superior mesenteric arteries (connected
isosurfaces), and skeletons superimposed on maximum intensity projections. All
algorithms were produced on a Unix-based workstation (Onyx; Silicon Graphics,
Mountain View, CA). The following visualization methods were generated.
|
|
|
|
|
|
|
|
|
|
|
|
Maximum intensity projection.—A projection image was obtained in a coronal plane that consisted of the points of maximum intensity along projected rays. The window and level of the maximum intensity projection were adjusted to accentuate the arteries.
Targeted maximum intensity projection.—Thin-section maximum intensity projections were generated to eliminate extraneous signal intensity from nonvascular tissue. Three targeted maximum intensity projections, each roughly encompassing a third of the coronal slices, were used. The window or level was adjusted separately for each targeted maximum intensity projection to emphasize the arteries.
Isosurfaces.—In this method of display, a mesh of triangles was first developed with an isosurface contour tool. The vertices were all located within points with the same interpolated image intensity. Lighting and shading were added to produce a 3D effect. Preset threshold values were used to generate isosurfaces at 20%, 25%, and 30% of the maximum voxel intensity in the image.
Connected isosurface.—A surface rendering was generated as previously described with the constraint that the isosurface must be in continuity with a seed point placed by the user. Seed points were placed over origins of the celiac and superior mesenteric arteries and propagated over the isosurface. This procedure reduced background scatter and improved the depiction of the vessels.
Skeletons.—To generate a skeleton [7], the user established the seed point and endpoint of each vessel on source images, and the skeleton algorithm determined the most likely path connecting those points and produced a skeleton of the vessel. Starting seed points were placed on the celiac and superior mesenteric artery origins on the source images. Endpoints were placed on the small distal vessels seen on source images. One author generated the skeletons. The connected paths followed the geometric centerlines of the vessels. The skeletons were superimposed on the maximum intensity projection for display purposes. Once the connectivity pattern was established, in approximately 4 min, the operator determined if an unsatisfactory path was produced because of either inaccurate specification of the endpoint, absence of the vascular path due to the limited field of view of the image, or other problems causing errors in the connectivity structure. An anomalous result was obvious to the operator because it created a chaotic path. The skeleton was transformed to a tapered tubular-surface mesh that was shaded (to provide depth cues), and each vessel was selectively colorized. The celiac artery and its branches were color-encoded red; the superior mesenteric artery was color-encoded blue to distinguish the two vessel trees. The user indicated the origin of the celiac and superior mesenteric artery. The shaded surface derived from the skeleton downstream from that point was then selectively colorized. The selection of the origin of the vessel for this purpose was done independently from the selection of the seed points of the skeleton.
Observer Study
Four observers who were unaware of the surgical results reviewed each
method of display in random order on a workstation. Each observer was asked to
identify (if present) the main hepatic artery, right hepatic artery, left
hepatic artery, replaced right hepatic artery, accessory left hepatic artery
arising from the left gastric artery, and the superior mesenteric artery. If
the vessel was not present, it was assigned a score of zero. If the vessel
could be identified, a confidence rating ranging from 1 to 5 (1 = not
confident, 5 = very confident) was assigned.
A receiver operating characteristic (ROC) curve was generated for each
method of displaying the hepatic arterial anatomy, and the area under the ROC
curve (Az) was calculated. The sensitivity and specificity
of each display method were determined with different thresholds for
positivity or negativity (i.e., 
1, 2). Additionally, interobserver
variability was determined for paired reviewers with kappa scores. The time
for each observer to interpret each method was recorded. Data were analyzed
with statistical software (Analyze-It version. 1.44; Analyze-It Software,
Leeds, UK).
|
TopAbstractIntroductionSubjects and MethodsResultsDiscussionReferences |
|---|
The ROC analysis (Table 1 and Fig. 3) showed superior performance for skeletons compared with maximum intensity projection, targeted maximum intensity projection, isosurface, and connected isosurface renderings. Skeletonized MR angiograms had the highest area under the ROC curve (Az, 0.90 ± 0.04) compared with the other techniques, (connected isosurface, 0.83 ± 0.05; targeted maximum intensity projection, 0.82 ± 0.05; maximum intensity projection, 0.79 ± 0.05; isosurface, 0.78 ± 0.06). Pairwise comparisons showed statistically significant differences between the performance of the skeleton and the next best technique, connected isosurfaces (p = 0.013). Pairwise comparisons between Az skeleton and the other techniques varied in significance from a p value of 0.001 to a p value of 0.0001 (Fig. 3).
|
|
Kappa scores of agreement were also highest for skeletonized MRA (0.75 ± 0.04) compared with the other techniques (maximum intensity projection, 0.58 ± 0.05; connected isosurface, 0.55 ± 0.05; isosurface, 0.42 ± 0.06; targeted maximum intensity projection, 0.37 ± 0.07). The 95% confidence intervals (CI) showed no overlap between kappa scores for skeletons and those of the other techniques (Fig. 4).
|
Compared with source images, all visualization methods were faster to interpret, but skeletons were more quickly interpreted than the other techniques (p = 0.04). The times to interpret for skeletons were 24.2 ± 3.8 sec; maximum intensity projection, 28.5 ± 4.0 sec; connected isosurface, 28.4 ± 3.4 sec; isosurface, 32.4 ± 3.3 sec; targeted maximum intensity projection, 39.2 ± 4.5 sec. The interpretation time was significantly different between skeletons and the next fastest technique (p = 0.04), connected isosurfaces, but the absolute magnitude of the difference was small (4.2 sec). As we predicted, interpreting the source images took the longest time with a mean time of 82 sec (95% CI, 72-108%). Image-generation times varied with the case. The maximum intensity projection took only about 10-20 sec, whereas the targeted maximum intensity projection could take an additional 1-5 min of user interaction. The pure isosurfaces varied from 2 to 8 min, and the connected isosurface took from 30 sec to 4 min. The skeleton took 6-10 min of processing.
An overall sensitivity and specificity for aberrant anatomy were determined for each technique by averaging each reviewer's response and by considering a score of 3 or higher as positive. The sensitivities and specificities, respectively, for aberrant anatomy for each of the techniques were as follows: maximum intensity projection, 56% and 92%; the targeted maximum intensity projection, 57% and 95%; isosurface, 62% and 89%; connected isosurface, 69% and 94%; and skeleton, 88% and 90%.
|
TopAbstractIntroductionSubjects and MethodsResultsDiscussionReferences |
|---|
Other methods have been proposed to improve the rendering of vessels from 3D data sets. Filtering techniques attempt to suppress background signal intensity to decrease the likelihood that the maximum intensity projection will detect nonvascular high signal intensity [8,9,10,11,12]. These methods show promise in improving depiction of anatomy but do not address problems related to high-signal-intensity overlapping structures that limit the vascular anatomy that is displayed. Moreover, they do not permit higher level visualization methods such as 3D surface shading and colorization. The skeleton algorithm generates skeletons by identifying vascular-tree patterns in contiguous pixels. Skeleton technique applies rule-based algorithms that follow the course of a vessel through its geometric centerline. Several previous methods of vessel tracking have encountered difficulties at bifurcations, but this problem does not occur in skeletons [13]. Skeletons do not truncate small vessels as severely as maximum-intensity-projection algorithms do.
Skeletonization also has limitations. It requires an informed user to identify the origins and endpoints of vessels. However, this task is no more difficult than evaluating source images and, in fact, could easily be integrated into a source-image review. Some bias may have been introduced in this study because one of the reviewers created the skeletons interpreted by the other reviewers, but the study does show that the technique produces less variability in interpretation than the other visualization methods. Additionally, the actual diameter of the vessel can now be depicted on the skeleton with the centerline as the backbone of the skeleton [13]. Another limitation is that if the signal-to-noise ratio of the original MR angiogram is poor or is highly degraded by motion, the algorithm may not "find" an accurate vascular pathway. This problem may also occur with occluded arterial segments. The technique is also moderately computer-intensive, but recent improvements in computing power make this disadvantage minor. For instance, we have implemented the algorithm on a desktop 500-MHz Windows NT computer (Optiplex; Dell, Austin, TX) that requires processing times of only about 2 min. Finally, the complexity of using a noncommercial image-processing algorithm may discourage its implementation outside research settings. To encourage its use, we will make the technique available at www.cc.nih.gov/drd/software.html.
MRA is a valuable noninvasive method of depicting the hepatic arterial tree [14,15,16]. Skeletonization improves the reproducibility, accuracy, and time for interpretation of hepatic MRA. Skeletonization could have broad applicability not only for revealing hepatic vessels but also for depicting other vascular distributions. Moreover, it should be directly applicable to other methods of 3D angiographic methods such as CT angiography [17].
In conclusion, skeletonization of MRA based on connectivity algorithms is a new method for displaying 3D MR angiographic data. Compared with other existing techniques, it improves reviewer confidence, decreases interobserver variability, and decreases the time of interpretation. Skeletonization is likely to be most useful in the display of small-and intermediate-sized vessels.
Acknowledgments
We thank Barbara McCoy for her assistance in preparing this manuscript.
|
TopAbstractIntroductionSubjects and MethodsResultsDiscussionReferences |
|---|
CiteULike 
Complore 
Connotea 
Del.icio.us 
Digg 
Reddit 
Technorati What's this?
This article has been cited by other articles:
|
|
 |
|
  J. K. Kim, J. H. Kim, S.-J. Bae, and K.-S. Cho CT Angiography for Evaluation of Living Renal Donors: Comparison of Four Reconstruction Methods Am. J. Roentgenol., August 1, 2004; 183(2): 471 - 477. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 B. F. Tomandl, N. C. Kostner, M. Schempershofe, W. J. Huk, C. Strauss, L. Anker, and P. Hastreiter CT Angiography of Intracranial Aneurysms: A Focus on Postprocessing RadioGraphics, May 1, 2004; 24(3): 637 - 655. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
B. F. Tomandl, P. Hastreiter, S. Iserhardt-Bauer, N. C. Kostner, M. Schempershofe, W. J. Huk, T. Ertl, C. Strauss, and J. Romstock Standardized Evaluation of CT Angiography with Remote Generation of 3D Video Sequences for the Detection of Intracranial Aneurysms RadioGraphics, March 1, 2003; 23(2): e12 - e12. [Abstract] [Full Text]
|
|
|
|
 |
|
 A. M. Covey, L. A. Brody, M. A. Maluccio, G. I. Getrajdman, and K. T. Brown Variant Hepatic Arterial Anatomy Revisited: Digital Subtraction Angiography Performed in 600 Patients Radiology, August 1, 2002; 224(2): 542 - 547. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
= skeleton, 
=
isosurface, 
= targeted maximum intensity projection, [UNK] = connected
isosurface, [UNK] = maximum intensity projection.