(Radiology. 2001;219:288-293.)
© RSNA, 2001
Reduction of Patient Motion Artifacts in Digital Subtraction Angiography: Evaluation of a Fast and Fully Automatic Technique1
Erik H. W. Meijering, PhD,
Wiro J. Niessen, PhD,
Jeannette Bakker, MD, PhD,
Aart J. van der Molen, MD,
Gerard A. P. de Kort, MD,
Rob T. H. Lo, MD,
Willem P. T. M. Mali, MD, PhD and
Max A. Viergever, PhD
1 From the Image Sciences Institute, University Medical Center Utrecht, the Netherlands. Received March 23, 2000; revision requested May 13; revision received July 31; accepted August 29. Address correspondence to E.H.W.M., Biomedical Imaging Group, Swiss Federal Institute of Technology, EPFL/DMT/IOA/BIG, BM-Ecublens, CH-1015 Lausanne, Switzerland (e-mail: erik.meijering @epfl.ch).
 |
ABSTRACT
|
|---|
The performance of an automatic technique for the reduction of patient motion artifacts in digital subtraction angiography was evaluated. Four observers assessed the quality of 104 cerebral digital subtraction angiographic images that were processed by means of both the automatic technique and manual pixel shifting. The automatic technique resulted in better image quality and was considerably less time-consuming.
Index terms: Cerebral angiography, technology, 10.12483 • Digital subtraction angiography, technology, 10.12483 • Images, artifact • Images, processing • Images, quality
|
INTRODUCTION
|
|---|
Patient motion artifacts are a major cause of image quality degradation in digital subtraction angiography (DSA). Although several techniques have been proposed (1) during the past 2 decades to improve the acquisition of DSA images in relation to this problem, motion artifacts cannot be entirely avoided. Currently, the only postprocessing techniques available with clinical DSA devices are manual remasking and pixel shifting, which allow for reduction of artifacts caused by uniform translational motion only (1,2). Generally, however, patient movements have a more complex nature that limits the effectiveness of these reduction techniques. This problem has been recognized by researchers in the field of image processing and has been the incentive for the development of a number of semiautomatic or even fully automatic nonlinear retrospective registration techniques (1). However, apart from two exceptions (3,4), clinical evaluations of these techniques have, to our knowledge, never been reported. Another major problem with these techniques is that they are too time-consuming for use in clinical practice.
Recently, a fully automatic registration technique was developed that is capable of nonlinearly aligning pairs of images within less than 1 second (5). In this study, we evaluated the effectiveness of the automatic technique in reducing patient motion artifacts by comparing it with manual pixel shifting. The study was performed by using cerebral DSA images.
|
Materials and Methods
|
|---|
Images and Equipment
During a 5-month period, 104 cerebral radiographic angiographic series from 21 patients (13 men, eight women; age range, 28–82 years; mean age, 53 years) were archived digitally. From each series, we randomly selected one mask-contrast image pair for which the corresponding DSA image had been printed on film by the radiologists at the time of the examination. All images had been acquired by using a C-arm imaging system (Integris V3000; Philips Medical Systems, Best, the Netherlands) with a 20- or 25-cm image intensifier, a matrix size of either 512 x 512 (31 images) or 1,024 x 1,024 pixels (73 images), and a gray-level resolution of 10 bits per pixel.
Postprocessing operations and image quality assessments were performed by using a workstation (Octane; Silicon Graphics, De Meern, the Netherlands) with one 195-MHz MIPS R10000 processor, 256-Mbyte of main memory (instruction and data cache size both 32 kbyte), and an IMPACTSR graphics board with 4-Mbyte texture memory. All images were displayed in a window of 700 x 700 pixels on a 19-inch monitor (Silicon Graphics), which had a resolution of 1,280 x 1,024 pixels (refresh rate, 75 Hz). By using this window, images were displayed with the same effective diameter (approximately 11.5 inches) as they are usually displayed on the 15-inch progressive display monitor of the imaging system. The contrast and brightness settings of the window were fixed during the evaluation.
Manual and Automatic Registration
Manual correction for motion artifacts on the DSA images corresponding to the 104 mask-contrast pairs was performed by using a special pixel-shifting tool that could be executed on the workstation and that exactly mimicked the pixel-shifting facility on the viewing console of the imaging system used in daily practice. Automatic correction was performed by using the algorithm by Meijering et al (5), briefly described next (see also Fig 1).
View larger version (112K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 1a. Example of the stages in the automatic registration algorithm evaluated in this study. (a) Mask image of a lateral cerebral DSA image. (b) Output of the edge-detection algorithm (gray regions) and control-point selection mechanism (white dots). (c) Live image overlaid with the automatically computed local displacement vectors indicating the correspondence with the mask image. From the vector field, it is clear that in this example the patient’s movement was of a rotational nature. (d) Triangulation of the set of control points used for interpolation of the displacement vectors and final warping of the mask image.
|
|
View larger version (38K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 1b. Example of the stages in the automatic registration algorithm evaluated in this study. (a) Mask image of a lateral cerebral DSA image. (b) Output of the edge-detection algorithm (gray regions) and control-point selection mechanism (white dots). (c) Live image overlaid with the automatically computed local displacement vectors indicating the correspondence with the mask image. From the vector field, it is clear that in this example the patient’s movement was of a rotational nature. (d) Triangulation of the set of control points used for interpolation of the displacement vectors and final warping of the mask image.
|
|
View larger version (117K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 1c. Example of the stages in the automatic registration algorithm evaluated in this study. (a) Mask image of a lateral cerebral DSA image. (b) Output of the edge-detection algorithm (gray regions) and control-point selection mechanism (white dots). (c) Live image overlaid with the automatically computed local displacement vectors indicating the correspondence with the mask image. From the vector field, it is clear that in this example the patient’s movement was of a rotational nature. (d) Triangulation of the set of control points used for interpolation of the displacement vectors and final warping of the mask image.
|
|
View larger version (154K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 1d. Example of the stages in the automatic registration algorithm evaluated in this study. (a) Mask image of a lateral cerebral DSA image. (b) Output of the edge-detection algorithm (gray regions) and control-point selection mechanism (white dots). (c) Live image overlaid with the automatically computed local displacement vectors indicating the correspondence with the mask image. From the vector field, it is clear that in this example the patient’s movement was of a rotational nature. (d) Triangulation of the set of control points used for interpolation of the displacement vectors and final warping of the mask image.
|
|
First, the algorithm applies edge detection to the mask image to extract regions that have a high potential for showing artifacts. Next, control points for the warping operation are automatically selected at local maxima of the gradient magnitude while constraining the minimum and maximum distance between these points. The local displacements of image structures at the control points are then computed by means of a template-matching procedure based on the energy similarity measure (6), followed by detection and correction of inconsistent displacement vectors by comparison with neighboring vectors. Finally, the mask image is warped according to the displacement vector field resulting from linear interpolation of the local displacements at the control points. This is performed efficiently by using triangulation of the control points with hardware-accelerated texture mapping. For the present study, all algorithm parameters were fixed to the values proposed in reference 5.
Method of Evaluation
Four observers, three radiologists (A.J.v.d.M., G.A.P.d.K., R.T.H.L.) and a resident (J.B.), participated in the evaluation, which consisted of two parts. In the first part, manual registrations of the 104 mask-contrast image pairs were performed separately and independently by the four observers by using the pixel-shifting tool. Since optimal manual registration of an image pair is task-dependent, the observers were provided with the clinical indication for the acquisition of the images, which was a cerebral aneurysm (41 images, seven patients), a stenosis in the carotid arteries (34 images, eight patients), a tumor (20 images, four patients), or vasculitis (nine images, two patients). However, the images were presented in random order. The final horizontal and vertical mask shifts for all mask-contrast pairs as indicated by each of the observers were stored automatically by the computer, with the time each observer required to perform the manual registration of each pair. The resulting manually corrected DSA images were also stored. The DSA images resulting from automatic registration of all mask-contrast pairs were computed and stored separately.
The second part of the study concerned the comparison of the quality of the automatically and manually corrected DSA images. To this end, the following three DSA image pairs were formed for each of the 104 original DSA images: (a) automatically corrected DSA image and original DSA image, (b) manually corrected DSA image and original DSA image, and (c) automatically corrected DSA image and manually corrected DSA image. This resulted in a total of 312 DSA image pairs, which were presented to the observers. Although the original and automatically corrected DSA images were the same for all four observers, each of the observers was presented with his or her own manual corrections resulting from the first part. For each of the DSA image pairs, the differences between the two images (denoted image A and image B) could be assessed by alternating the image that was displayed.
The observers were given the clinical information for all images and were asked to rate the relative quality of the two images by choosing one of the following: AB, image A and image B were similar (ie, the number of artifacts and the magnitude of the artifacts were the same in the diagnostically relevant parts or on the entire images); A+, image A was better than image B; A++, image A was much better than image B (ie, the number of artifacts or the magnitude of the artifacts on image A was smaller [A+] or much smaller [A++] than that of image B in the diagnostically relevant parts or on the entire image); B+, image B was better than image A; or B++, image B was much better than image A (ie, the number of artifacts or the magnitude of the artifacts on image B was smaller [B+] or much smaller [B++] than that of image A in the diagnostically relevant parts or on the entire image).
Similar to the first part, the second part of the study was performed separately and independently by the four observers. However, prior to the second part, there was a meeting among the observers to obtain consensus regarding the rating of relative image quality. For this consensus meeting, 10 sample cerebral DSA image pairs that were not included in the actual study were used.
To avoid bias in the ratings, the images were presented to the observers in a completely randomized and blinded fashion; not only were the 312 DSA image pairs randomized, but also the order of the images within each pair was randomized, and the observers were not informed of the type of correction (no, manual, or automatic correction) that was applied to the images. Furthermore, to reduce the possibility of observers recognizing their own manual corrections, the period between the first and the second parts of the study was at least 3 weeks for each of the observers.
Statistical Analyses
Interobserver agreement for the image quality ratings resulting from the second part of the study was assessed by using a 
test. To account for the degree of disagreement, we used the weighted w test; the weights for discrepancies of zero, one, two, three, and four categories in the ratings were 1, 0.75, 0.5, 0.25, and 0, respectively (7,8). Six w values were computed on the basis of a comparison of the ratings of two observers at a time. A w value of 1.0 indicates that the agreement is perfect, and a value of 0 indicates that it is not different from chance agreement. For the interpretation of w values in between these extremes, we used the Landis-Koch guidelines (8).
The ratings resulting from the second part allowed us to make both implicit and explicit comparisons of the effectiveness of the automatic and the manual registration technique in reducing motion artifacts. For this purpose, the ratings of the 312 DSA image pairs were divided into three groups: (a) ratings expressing the quality of automatically corrected DSA images relative to corresponding original (uncorrected) DSA images (or vice versa), (b) ratings expressing the quality of manually corrected DSA images relative to corresponding original (uncorrected) DSA images (or vice versa), and (c) ratings expressing the quality of automatically corrected DSA images relative to corresponding manually corrected DSA images (or vice versa).
Since the images of each pair were presented in random order, the original ratings were converted by using the rules presented in Table 1 to express the quality of any of the images in a given pair in terms of the other. Implicit comparison of the performance of the automatic and manual registration technique was then obtained by constructing a frequency table of the converted ratings from groups a and b. Explicit comparison was obtained by analyzing the ratings from group c. These comparisons were performed separately for the results of each observer. A comparison based on the mean frequencies was also performed.
The statistical significance of the possible improvement of the automatic registration technique compared with manual pixel shifting was assessed by using a 
2 test (7,8) applied to the frequency tables containing the ratings from groups a and b. Since one of the variables in these tables (viz, relative image quality) represents ordered categories, we did not use the ordinary 2 test but the more powerful 2 test for linear trend (7,8). For this test, we used uniform spacing of the categories. The null hypothesis was that the automatic and manual registration technique would be equally effective in reducing motion artifacts. A probability of P less than .05 for this hypothesis was chosen to indicate a statistically significant difference between the two techniques.
|
Results
|
|---|
In the first part of the study, the observers found that most of the 104 original cerebral DSA images could be improved to some extent by means of manual correction since, on average, 92 (88%) of 104 mask shift parameters differed from zero. The maximum shift recorded in either direction was 8.0 pixels, while the mean length of the shift vectors of all four observers was 1.2 pixels. From this it may be concluded that, although in some cases patient motion was severe, in most cases the motion artifacts were due to only relatively small displacements. The timing information stored along with the shift parameters revealed that manual correction required, on average, about 12 seconds (median, 7 seconds) per DSA image. In contrast, the automatic registration algorithm required, on average, only about 1 second (median, 1 second) per DSA image.
The w values computed from the ratings in the second part ranged from 0.60 to 0.71. According to the Landis-Koch guidelines, this indicates substantial agreement. Therefore, we restrict ourselves to presenting the means. The mean frequencies of the ratings from groups a and b, as described in Statistical Analyses, are presented in the first two rows in Table 2. From the implicit comparison based on the results in Table 2, it follows that, overall, the automatic registration technique resulted in better image quality compared with manual pixel shifting. To apply the 2 test for trend, the frequencies in the double negative (much worse) and negative (worse) columns had to be combined, since this test does not allow rows or columns to be entirely filled with zeros. The 2 test for trend applied to the modified frequency table showed that the probability for the null hypothesis of equal effectiveness to be true was P less than .05; from this result, it can be concluded that the automatic correction technique was significantly better than manual pixel shifting in reducing motion artifacts. We note that the same conclusion was found when this test was applied to the results of the separate observers. The mean frequencies of the ratings from group c, representing the results of the explicit comparison of the quality of automatically and manually corrected images, are presented in the last row in Table 2. These results support the conclusion drawn from the implicit comparison.
Two examples of cases in which the automatic registration technique was found to be superior to manual pixel shifting are shown in Figures 2 and 3, where the artifacts on the original DSA images are located primarily in the lower part of the image, around the main vessels. Although the artifacts could be removed to some extent by means of manual pixel shifting, it was not possible to completely remove them by using this technique. In the case of Figure 2, pixel shifting even resulted in a deterioration of artifacts in the lower right part of the image. Application of the automatic registration technique, on the other hand, resulted in overall correction and improved vessel visibility.
View larger version (132K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 2a. Example of a case in which the automatic registration technique was superior to manual pixel shifting. (a) Original lateral cerebral DSA image. (b, c) Resulting DSA images after manual registration by means of pixel shifting for two of the four observers. Because of the rotational nature of the patient’s movement, it was not possible to obtain an overall optimal correction of motion artifacts by means of this technique. From these two images, it is clear that a reduction of artifacts in one part of the image (in this example, the lower left parts) may result in a deterioration of artifacts elsewhere (in this example, the lower right parts). (d) DSA image resulting from application of the automatic registration technique.
|
|
View larger version (135K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 2b. Example of a case in which the automatic registration technique was superior to manual pixel shifting. (a) Original lateral cerebral DSA image. (b, c) Resulting DSA images after manual registration by means of pixel shifting for two of the four observers. Because of the rotational nature of the patient’s movement, it was not possible to obtain an overall optimal correction of motion artifacts by means of this technique. From these two images, it is clear that a reduction of artifacts in one part of the image (in this example, the lower left parts) may result in a deterioration of artifacts elsewhere (in this example, the lower right parts). (d) DSA image resulting from application of the automatic registration technique.
|
|
View larger version (137K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 2c. Example of a case in which the automatic registration technique was superior to manual pixel shifting. (a) Original lateral cerebral DSA image. (b, c) Resulting DSA images after manual registration by means of pixel shifting for two of the four observers. Because of the rotational nature of the patient’s movement, it was not possible to obtain an overall optimal correction of motion artifacts by means of this technique. From these two images, it is clear that a reduction of artifacts in one part of the image (in this example, the lower left parts) may result in a deterioration of artifacts elsewhere (in this example, the lower right parts). (d) DSA image resulting from application of the automatic registration technique.
|
|
View larger version (117K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 2d. Example of a case in which the automatic registration technique was superior to manual pixel shifting. (a) Original lateral cerebral DSA image. (b, c) Resulting DSA images after manual registration by means of pixel shifting for two of the four observers. Because of the rotational nature of the patient’s movement, it was not possible to obtain an overall optimal correction of motion artifacts by means of this technique. From these two images, it is clear that a reduction of artifacts in one part of the image (in this example, the lower left parts) may result in a deterioration of artifacts elsewhere (in this example, the lower right parts). (d) DSA image resulting from application of the automatic registration technique.
|
|
View larger version (126K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 3a. Another example of a case in which the automatic registration technique was superior to pixel shifting. (a) Original tilted oblique frontal cerebral DSA image. (b, c) Resulting DSA images after manual registration by means of pixel shifting, for two of the four observers. Similar to Figure 2, the patient’s movement as projected in the imaging plane was more complex than uniform translation. Although the artifacts were reduced substantially (lower parts of the images), it was not possible to obtain an overall optimal correction of motion artifacts by means of this technique. (d) DSA image after application of the automatic registration technique, resulting in improved visibility of the main vessel.
|
|
View larger version (115K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 3b. Another example of a case in which the automatic registration technique was superior to pixel shifting. (a) Original tilted oblique frontal cerebral DSA image. (b, c) Resulting DSA images after manual registration by means of pixel shifting, for two of the four observers. Similar to Figure 2, the patient’s movement as projected in the imaging plane was more complex than uniform translation. Although the artifacts were reduced substantially (lower parts of the images), it was not possible to obtain an overall optimal correction of motion artifacts by means of this technique. (d) DSA image after application of the automatic registration technique, resulting in improved visibility of the main vessel.
|
|
View larger version (120K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 3c. Another example of a case in which the automatic registration technique was superior to pixel shifting. (a) Original tilted oblique frontal cerebral DSA image. (b, c) Resulting DSA images after manual registration by means of pixel shifting, for two of the four observers. Similar to Figure 2, the patient’s movement as projected in the imaging plane was more complex than uniform translation. Although the artifacts were reduced substantially (lower parts of the images), it was not possible to obtain an overall optimal correction of motion artifacts by means of this technique. (d) DSA image after application of the automatic registration technique, resulting in improved visibility of the main vessel.
|
|
View larger version (109K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 3d. Another example of a case in which the automatic registration technique was superior to pixel shifting. (a) Original tilted oblique frontal cerebral DSA image. (b, c) Resulting DSA images after manual registration by means of pixel shifting, for two of the four observers. Similar to Figure 2, the patient’s movement as projected in the imaging plane was more complex than uniform translation. Although the artifacts were reduced substantially (lower parts of the images), it was not possible to obtain an overall optimal correction of motion artifacts by means of this technique. (d) DSA image after application of the automatic registration technique, resulting in improved visibility of the main vessel.
|
|
|
Discussion
|
|---|
Of the 104 DSA images included in this study, 38 (37%) had been manually corrected by means of pixel shifting before being printed on film and stored in the archive of our hospital. From the fact that, in the first part of the study, the observers found that no less than 92 (88%) of the images could be improved to some extent by using this technique, we conclude that in practice more images contain motion artifacts than are usually corrected. Manual correction of all images by means of pixel shifting is a labor-intensive operation: The results of our study indicated that, on average, 12 seconds per DSA image were required to optimally apply this technique. Apart from resulting in better overall image quality compared with manual pixel shifting, the automatic technique was considerably faster. On average, it required only 1 second per DSA image. Moreover, the technique did not require any effort from the radiologist.
We note that although the w test indicated substantial agreement between the observers, there was some spread in the individual w values. This may be due to the fact that in the second part of the study, the observers were presented with their own manual corrections, which were sometimes different for the different observers. Furthermore, although the results of the 2 test for trend—applied to either the mean results or the results of the individual observers—indicated a statistically significant difference (P < .05) in image quality after application of the automatic correction technique and manual pixel shifting, the outcome of an ordinary 2 test would have been somewhat less persuasive: For two of the observers, this test would have resulted in a P value of less than .1 for the null hypothesis to be true. However, the 2 test for the trend was more appropriate, since we were dealing with ordered categories.
As clearly illustrated by the examples in Figures 2 and 3, manual pixel shifting often results in improved image quality in and near the diagnostically relevant parts of an image, but it may sometimes result in deterioration of artifacts in other parts. This deterioration was a direct consequence of the fact that, with this technique, patient motion as projected in the imaging plane is assumed to be uniformly translational. One may argue that this is not really a problem in practice as long as artifacts can be reduced in the diagnostically relevant parts of the image and that it is therefore sufficient to use manual pixel shifting rather than a more sophisticated automatic correction technique. However, even if there were no difference in performance from the point of view of image quality, it would still be advantageous to use the automatic technique evaluated in this study, since it was considerably less time-consuming than manual pixel shifting.
The fact that the automatic registration technique performed significantly better than manual pixel shifting does not imply that in practice the former technique will always be better than the latter. The mean results from the explicit comparison of the two techniques (last row in Table 2) indicated that in 5% of all cases, the corrected DSA image resulting from manual pixel shifting was better than the corresponding automatically corrected DSA image. We observed that in these cases the automatic technique did not introduce new artifacts. However, it was unable to reduce some of the artifacts at the borders of the images. This may be caused by the lack of image content in those regions, which reduces the possibilities for any template matching procedure to find the correct local displacement vectors.
In most reports on the reduction of motion artifacts on DSA images, the evaluation of newly developed techniques involved only one or at most a few clinical DSA images or phantoms, and the quality of the resulting corrected images was assessed by the same persons that developed the algorithm. To our knowledge, the only more elaborate and objective evaluation studies are the ones by Takahashi et al (3) and Hayashi et al (4). In the former study, three techniques were evaluated: manual remasking, manual pixel shifting, and an automatic registration technique. It was concluded that remasking was most effective. It was also found that, after remasking was applied, remaining artifacts were reduced equally well by means of manual pixel shifting and their automatic registration technique. In the study by Hayashi et al (4), the authors compared two techniques: manual pixel shifting and an automatic registration technique developed by some of the co-authors of that study. In 14 of 16 cerebral DSA image series, the images resulting from the automatic registration technique had better quality. In the other two, the techniques performed comparably.
Because of the lack of detailed information provided by the authors of these studies, it is difficult to explicitly compare their findings to ours. We note, however, that the automatic registration technique evaluated in our study was based on approaches that have been shown to yield faster and more accurate registrations compared with their techniques (see the articles by Meijering et al [1,5] for more technical details). Hayashi et al (4) reported that their algorithm requires about 8 minutes of computation time. In contrast, the algorithm evaluated in our study requires, on average, about 1 second per DSA image, which certainly makes it more suitable for use in clinical practice. Furthermore, owing to the use of better similarity measures in the template matching procedure, the images processed by our algorithm are comparable to, better than, or even much better than those resulting from manual pixel shifting in 95% of all cases.
Finally, our study involved only images that were already considered clinically useful. Frequently, during acquisition, the patient’s movements are too severe to result in diagnostically useful DSA images, even when pixel shifting is used afterward. In such cases, the series is repeated. In some cases, the automatic registration technique might help prevent a second DSA series. Online availability of the automatically corrected DSA images would offer the radiologist the possibility to directly check whether a new series of images must be acquired and, thereby, avoid the need to go back to the console to check it manually by means of pixel shifting. We also note that, in our study, we were interested only in overall image quality improvement, without relation to specific diagnostic tasks, such as the grading of stenoses or the detection of small aneurysms. It may be that the automatic registration technique also provides an improvement in that respect compared with manual pixel shifting. Confirmation of these claims is the goal of future studies.
|
ACKNOWLEDGMENTS
|
|---|
The authors are grateful to Philips Medical Systems, Best, the Netherlands, for making available their clinical Octane workstation, on which the evaluation was performed. Gerard van Hoorn, MSc, Tineke Kievit, Wilma Pauw, Koen L. Vincken, PhD, Theo van Walsum, PhD, Remko van der Weide, MSc, and Onno Wink, MSc (Image Sciences Institute, Utrecht, the Netherlands) are acknowledged for their assistance in the start-up phase of this study.
|
FOOTNOTES
|
|---|
Abbreviation: DSA = digital subtraction angiography
Author contributions: Guarantor of integrity of entire study, E.H.W.M.; study concepts, all authors; study design, E.H.W.M., W.J.N.; literature research, E.H.W.M.; experimental studies, J.B., A.J.v.d.M., G.A.P.d.K., R.T.H.L.; data acquisition, all authors; data analysis/interpretation, E.H.W.M.; statistical analysis, E.H.W.M.; manuscript preparation and editing, E.H.W.M.; manuscript definition of intellectual content, revision/review and final version approval, all authors.
|
REFERENCES
|
|---|
-
Meijering EHW, Niessen WJ, Viergever MA. Retrospective motion correction in digital subtraction angiography: a review. IEEE Trans Med Imaging 1999; 18:2-21.[Medline]
-
Levin DC, Shapiro RM, Boxt LM, Dunham L, Harrington DP, Ergun DL. Digital subtraction angiography: principles and pitfalls of image enhancement techniques. AJR Am J Roentgenol 1984; 143:447-454.[Abstract/Free Full Text]
-
Takahashi M, Shinzato J, Korogi Y, et al. Automatic reregistration for correction of localized misregistration artifacts in digital subtraction angiography of the head and neck. Acta Radiol Suppl 1986; 369:281-284.[Medline]
-
Hayashi N, Sakai T, Kitagawa M, et al. Nonlinear geometric warping of the mask image: a new method for reducing misregistration artifacts in digital subtraction angiography. Cardiovasc Intervent Radiol 1998; 21:138-141.[Medline]
-
Meijering EHW, Zuiderveld KJ, Viergever MA. Image registration for digital subtraction angiography. Int J Comput Vision 1999; 31:227-246.
-
Buzug TM, Weese J. Image registration for DSA quality enhancement. Comput Med Imaging Graph 1998; 22:103-113.[Medline]
-
Fleiss JL. Statistical methods for rates and proportions 2nd ed. New York, NY: Wiley, 1981.
-
Altman DG. Practical statistics for medical research London, England: Chapman & Hall, 1991.
This article has been cited by other articles:
|
|
|
|
 
J. W. Olin, J. A. Kaufman, D. A. Bluemke, R. O. Bonow, M. D. Gerhard, M. R. Jaff, G. D. Rubin, and W. Hall
Atherosclerotic Vascular Disease Conference: Writing Group IV: Imaging
Circulation,
June 1, 2004;
109(21):
2626 - 2633.
[Full Text]
[PDF]
|
|
|
|
|
|
|
H. Schmitt, M. Grass, S. Heiland, S. Hahnel, and K. Sartor
Image Stacking with Entropy Values in Conventional Angiography: Initial Experience
Radiology,
January 1, 2004;
230(1):
294 - 298.
[Abstract]
[Full Text]
[PDF]
|
|
|
TOP
ABSTRACT
INTRODUCTION
