The purpose of this study was to im-prove dynamic two-dimensional pro-jection magnetic resonance digitalsubtraction angiography by using re-masking and filtering postprocessingtechniques. Four methods were eval-uated in 50 patients: default masksubtraction, remasked subtraction,filtering based on the SD, and linearfiltering. The results demonstratedthat postprocessing techniques suchas linear filtering can reduce back-ground motion artifacts and improvearterial contrast-to-noise ratio.
Magnetic resonance (MR) angiography isincreasingly used to help diagnose pe-ripheral vascular disease, particularly inpatients with contraindications to iodin-ated contrast material or arterial catheter-ization (1–6). Contrast material–enhancedMR angiography with the three-dimen-sional bolus-chase technique (6–13) hasbeen particularly useful because it reducesthe imaging time to a few minutes to coverthe entire peripheral vasculature. It alsoreduces the dose of contrast material re-quired because multiple stations are ac-quired during a single injection of MRcontrast material, which is analogous tobolus-chase conventional digital subtrac-tion angiography (DSA). However, de-spite recent developments in gradientspeed and k-space sampling, accuracy islimited below the knees where the arter-ies are smaller in caliber and the bolussynchronization is less reliable as timingis optimized for the pelvis (13).
For this reason, we have found it usefulto supplement the bolus-chase examina-tion with preliminary two-dimensional(2D) MR DSA of the trifurcation and feet(13–16). It requires only 5–7 mL of gado-linium-based contrast material for the tri-
furcation and thus does not significantlyinterfere with performing a three-dimen-sional bolus-chase examination subse-quently. Bolus timing information from2D MR DSA is also helpful in planningthe subsequent three-dimensional bolus-chase examination. Projection MR an-giography of the trifurcation providesonly a single anteroposterior view, but itcan be performed repeatedly at 1–2-sec-ond intervals to help evaluate the timecourse and symmetry of vascular en-hancement. It involves acquiring at leastone projection image prior to the arrivalof the contrast material to use as a maskfor subtraction of the background signal.In these ways, it is similar to conven-tional DSA.
In conventional DSA, postprocessingtechniques including remasking, filter-ing, and pixel shifting are commonlyused to improve image quality. It wouldbe valuable to assess the benefits of thesetechniques for MR DSA. Because MR dataare complex (vector), unlike radiographicdata, which are real numbers, these post-processing techniques have to be gener-alized to the complex domain. The pur-pose of this study was to evaluateremasking and filtering postprocessingtechniques for time-resolved MR DSA.
Materials and Methods
Patients
Fourier data from 2D MR DSA of thetrifurcation in 50 consecutive patientsundergoing peripheral MR angiographyfrom September 11 to November 25,2000, were evaluated with and withoutaveraging. This study was approved byour institutional review board; informedconsent was not required. These patientsincluded 29 men (age range, 24–87years; mean age, 70 years) and 21 women(age range, 33–85 years; mean age, 68years). The primary indications for pe-
564
ripheral MR angiography in these pa-tients included claudication (n � 26),limb-threatening ischemia (n � 13), an-eurysm (n � 7), bypass graft placement(n � 3), and dissection (n � 1).
Imaging
All data were obtained at 1.5 T by usingthe head coil for signal transmission andreception (LX Horizon; GE Medical Sys-tems, Milwaukee, Wis). The patients wereplaced feet first into the magnet, withtheir legs positioned within the head coilto image from above the patella down tothe middle of the calf. A sagittal gradient-echo scout sequence was used to positionthe coronal 2D projection MR angio-graphic slab so that it encompassed theentire calf. Two-dimensional projectionMR angiography was performed as acoronal spoiled gradient-echo sequenceby using the following parameters: 10/2(repetition time msec/echo time msec);flip angle, 60°; slab thickness, 7–10 cm;field of view, 30 cm; matrix, 256 � 192;bandwidth, 16 kHz. The imaging time
was 1.95 seconds per acquisition for atotal of 68 seconds to repeat the acquisi-tion 35 times.
Five to seven milliliters of 0.5 mol/Lgadolinium-based contrast material (gado-pentetate dimeglumine, Magnevist, Ber-lex Laboratories, Wayne, NJ or gadodi-amide, Omniscan, Nycomed Amersham,Princeton, NJ) was injected and flushedwith 20 mL of saline. The injection ratewas 2.5 mL/sec and was performed withan automatic injector or by hand. Theinjection was initiated simultaneouslywith initiating the imaging. In this way,at least five to 10 mask images were ob-tained before administration of the con-trast material and prior to the arrival inthe arterial phase at the trifurcation.
Postprocessing Techniques
Our approach to generalize postpro-cessing techniques to the complex do-main was to treat the two orthogonalcomponents of MR data separately andequivalently, apply postprocessing tech-niques of real numbers to each compo-nent, and combine the two processedcomponents. In this study, we focusedon remasking and filtering methods.
Remasking.—The purpose was to iden-tify a mask that was closest to the back-ground of the enhanced peak arterialphase image so that motion artifacts wereminimized on the subtracted arterial phaseimages. This was achieved through two it-erations of subtraction. First, subtractionby using a default mask (image 5 of the 35serial images) was performed to identifythe peak arterial phase (maximal arterialsignal). Then the identified peak arterialphase image was used as a mask, and sub-traction was performed again to identifythe optimal mask among the images ob-tained before contrast enhancement (ie,the one with minimal background).
Filtering.—A widely used filtering tech-nique is matched filtering, which sums atime series of images into a single imageof optimal contrast-to-noise ratio (CNR)with undesired background signal sup-pressed (17–19). The linearly matched fil-ter requires an input of signal waveformthat characterizes the arterial signal.However, such an arterial waveform isnot well defined throughout the field ofview because of variations in the time thecontrast material reaches the arteries,particularly portions of an artery distal to
Note.—Default � peak arterial phase imagefrom the default mask subtraction, linear �linearly filtered image, remasked � peak arte-rial phase image from the remasked subtrac-tion, SD � SD image.
TABLE 3Summary of Image Quality Comparison Results on the Five-Point Scale andCorresponding Statistical Significance
Image QualityDistribution
ScaleP Value
forImageQuality�2 �1 0 1 2
Reader 1Remasked versus
default0 4 33 11 2 �.3
SD versusdefault
0 2 5 31 12 �.003
SD versusremasked
0 2 12 32 3 �.003
Linear versusdefault
0 0 1 25 24 �.003
Linear versusremasked
0 0 0 27 23 �.003
Linear versus SD 0 9 19 20 2 �.06Reader 2
Remasked versusdefault
0 5 31 10 4 �.15
SD versusdefault
3 0 1 27 19 �.003
SD versusremasked
0 1 1 29 19 �.003
Linear versusdefault
0 0 1 28 21 �.003
Linear versusremasked
0 0 1 32 17 �.003
Linear versus SD 2 5 12 28 3 �.003
Note.—Default � peak arterial phase image from the default mask subtraction, linear � linearlyfiltered image, remasked � peak arterial phase image from the remasked subtraction, SD � SDimage. �2 � one image was substantially worse than the other, �1 � one image was modestlyworse than the other, 0 � one image was approximately the same as the other, 1 � one image wasmodestly better than the other, 2 � one image was substantially better than the other.
Volume 222 � Number 2 Postprocessing Techniques for Time-resolved Contrast-enhanced MR Angiography � 565
stenosis, small branch arteries, and arter-ies filled with retrograde flow.
This problem was overcome with thegeneralized local matched filter, whichreplaced the global arterial waveformwith a local arterial curve defined by thepixel value minus its mean. This providessimple automated filtering. Let the finalimage be s, the time series Sn, and itsaverage S, and then the local matchedfiltered image is calculated as follows,which is equivalent to taking the SD (19):
�Re�s�x, y�2
� �n
�Re�Sn�x, y � Re�S��x, y�2/N
�Im�s�x, y�2
� �n
�Im�Sn�x, y � Im�S��x, y�2/N
�s� x, y�
� ��Re�s�x, y�2 � �Im�s�x, y�2 1/ 2,
where Re � real part, x and y � in-planecoordinates, n � image index, Im �imaginary part, and N � the total num-ber of images in the series. This filteredimage is referred to as the SD image. Inthis study, the summation range was lim-ited to exclude images with venous orbackground enhancement.
On this semiautomated SD image, allimages were used from the first mask tothe last arterial phase, and it was prone toartifacts caused by motion that occurredduring imaging. Such motion artifactsare common at peripheral MR angiogra-phy, as the legs tremble in most olderpatients. As a first attempt to eliminatemotion artifacts, simple linear filtering ofa manually selected mask image set (M)and arterial phase image set (A) was usedto generate a linearly filtered image:
sh� x, y � �n�A
Sn� x, y/NA
� �n�M
Sn�x, y/NM .
Image Evaluation
The peak arterial phase image from thedefault mask subtraction (default) andfrom the remasked subtraction (re-masked), the SD image (SD), and the lin-early filtered image (linear) were evalu-ated by using an objective CNR and asubjective image quality score.
CNR was measured in the followingmanner. The vessel signal intensity wasthe average of the maximal signal inten-sity of a transverse line over all transverselines on an image. Noise was the SD ofthe signal intensity in a background re-
gion. The CNR was the difference be-tween vessel signal intensity and themean of background signal intensity di-vided by the SD of the latter. In addition,the first and second order branches of thegeniculate, anterior tibial, posterior tib-ial, and peroneal arteries were also iden-tified independently by two experiencedvascular radiologists (M.R.P., P.A.W.).
Paired comparisons of the four postpro-cessed images—default, remasked, SD, andlinear—for a total of six pairs, were per-formed by the same two MR radiologists toassess the differences in image quality byusing a five-point scale: 2, one image wassubstantially better than the other; 1, oneimage was modestly better than the other;0, one image was approximately the sameas the other; �1, one image was modestlyworse than the other; �2, one image wassubstantially worse than the other. The im-age pairs were presented randomly, andthe two radiologists read the images inde-pendently.
Statistical Analysis
The significance of differences in apaired comparison was assessed by per-forming the paired two-sample t test byusing the two sets of CNR measurements
and by performing the paired Wilcoxonsigned rank test over the image qualitydifference score set (20). The Bonferronicorrection was applied to the P value toaccount for the nature of multiple com-parisons (21).
Results
The CNR data are tabulated in Table 1,and the comparisons are tabulated in Ta-ble 2. The average CNRs for the peak ar-
TABLE 4Summary of Arteries Visualized by the Readers
Note.—Default � peak arterial phase image from the default mask subtraction, linear � linearlyfiltered image, remasked � peak arterial phase image from the remasked subtraction, SD � SDimage.
TABLE 5Summary of Image ArtifactOccurrence
Image
No. of Artifacts Observed
Reader 1 Reader 2
Default 23 22Remasked 15 16SD 14 14Linear 14 15
Note.—Default � peak arterial phase imagefrom the default mask subtraction, linear �linearly filtered image, remasked � peak arte-rial phase image from the remasked subtrac-tion, SD � SD image.
566 � Radiology � February 2002 Yoo et al
terial phase images from default subtrac-tion (default) and remasked subtraction(remasked) were approximately the same(P � .9). Both the SD image (SD) and thelinearly filtered image (linear) had signif-icantly higher (greater than twofold)CNR than did the default and remaskedimages (P � .001), as many images weresummarized. The linearly filtered imagealso had significantly lower CNR thandid the SD image (P � .001), as fewerimages were used in the linear filtering.
In the image quality comparison (Ta-ble 3), both readers 1 and 2 regarded thelinearly filtered and SD images to be sig-nificantly better than both the defaultand remasked images (P � .003). The lin-early filtered images were judged to bebetter than the SD images (P � .06 forreader 1 and P � .003 for reader 2). Re-masked images yielded slightly but notsignificantly better image quality thandid the default mask images according toboth readers (P � .3 for reader 1 and P �.15 for reader 2).
Vessel visibility was higher on the SDand linearly filtered images than on ei-ther the default or remasked images (Ta-
ble 4). For linearly and SD filtered images,the first order branch vessels of the genic-ulate, anterior tibial, posterior tibial, andperoneal arteries were seen at least twiceas often as they were seen on either thedefault or remasked images. There wereinstances in which the second orderbranches of these arteries were observedon the linear or SD images but never onthe default or remasked images. Differ-ences were less discernable in vessel visi-bility between SD and linear images. SDand linearly filtered images displayed lessmotion artifact than did both remaskedand default mask images (Table 5), whileremasked images displayed less motionartifact than did the default mask images.
Figure 1 demonstrates an instance inwhich the remasked image (Fig 1, D) elim-inated much of the background motionartifact on the arterial phase images createdby using the default mask (Fig 1, A–C). Ves-sel depiction improved on the linearly fil-tered image (Fig 1, F), as compared withthat on any of the individual arterial phaseimages. Linear filtering showed second or-der branches off the geniculate and poste-rior tibial arteries, whereas second order
branches off these arteries were not seenon either default or remasked images (Fig1, A–D). The SD image (Fig 1, E) also dem-onstrated improved vessel definition andshowed second order branches off thegeniculate and posterior tibial arteries,though motion artifact remained.
Figure 2 demonstrates the advantage ofmanually selecting an optimized maskover using an arbitrary default mask. Themotion artifact seen in the upper portionof the arterial phase images of defaultmask (Fig 2, A) was substantially reducedby remasking (Fig 2, B), which also madethe geniculate branches more visible. Thealmost automated SD image (Fig 2, C)contained motion artifacts, which weresubstantially reduced on the linearly fil-tered image (Fig 2, D).
Figure 3 is an example in which imagesin a patient with slow arterial blood flowbenefit from the single linearly filtered im-age. Individual default arterial phase im-ages for this patient yielded poor CNR andvessel visibility (Fig 3, A). However, the SDand linearly filtered images (Fig 3, C, D)showed increased visibility of the genicu-late artery and branches off the posteriortibial artery in comparison with that onthe default and remasked images (Fig 3, A,
Figure 1. Coronal 2D MR DSA (10/2; flip angle, 60°) images ob-tained in a 75-year-old man with claudication. A–C, Contiguousarterial phase images in which the default mask was used. D, Peakarterial phase image (corresponding to B) in which an optimizedmask was used. E, Filtered image with SD from the first mask to thelast arterial phase images. F, Linearly filtered image in which manu-ally selected masks (five nonenhanced images) and arterial phases(three arterial phases depicted in A–C) were used. The remasked image(D) demonstrates substantial reduction in background motion arti-facts, as compared with that on the default image (B). Both the SDimage (E) and the linearly filtered image (F) demonstrate improvedCNR and better vascular details, but the background artifacts aresuppressed on the linear image (F).
Figure 2. Coronal 2D MR DSA (10/2; flip angle, 60°) imagesobtained in a 67-year-old man with right calf claudication. A,Arterial phase image in which the default mask was used. B, Peakarterial phase image (corresponding to A) in which an optimizedmask was used. C, Filtered image with SD from the first mask tothe last arterial phase images. D, Linearly filtered image in whichmanually selected masks (10 nonenhanced images) and arterialphases (four arterial phases) were used. The motion artifact issubstantially suppressed on the remasked image (B) and thelinearly filtered image (D).
Volume 222 � Number 2 Postprocessing Techniques for Time-resolved Contrast-enhanced MR Angiography � 567
B). The CNRs of the linearly filtered and SDimages were more than twofold greaterthan those of the default mask or remaskedimages.
Discussion
These data in 50 consecutive patientsundergoing peripheral MR angiographydemonstrate that postprocessing can im-prove 2D MR DSA image quality, in con-cordance with the experience of radio-graphic DSA. Remasking can reducemotion artifacts on peak arterial phaseimages. Filtered images can summarilydepict vascular anatomy in the time se-ries images conveniently on a single im-age with significantly higher CNR.
The ability to have a single summaryimage reflecting the best depiction of vas-cular anatomy may reduce the time re-quired to analyze a case by having to lookat only a single image. Of course, the arte-rial phase images could be displayed as avideo loop and provide equivalent or moreinformation to the observer than would asummary image. Presumably, a human can
perform averaging equivalent to linear orSD filtering by integrating the informationof multiple images projected rapidly in se-quence, and this might be superior to us-ing the computer to generate a single com-posite image. This would be especially truein the event of motion that results in vesselblur when misregistered images are aver-aged. However, the summary image doesnot diminish but adds to the value of thetime series images. Even the video loopdisplay method would benefit from filter-ing by using a mask of reduced noise ob-tained by averaging multiple masks. Fur-thermore, there are many instances inwhich video is not available—for example,on the view box in an operating room.
The semiautomated local matched fil-tered image (SD image) demonstrates ar-terial anatomy superior to that on theindividual peak arterial phase image butinferior to that on the linearly filteredimage. This may be attributed to motionartifacts. Individual mask and arterialphase images with motion effects, whichwere included on the SD image but ex-cluded on the linearly filtered image,though improving CNR, introduce mo-tion noise that diminishes the conspicu-ity of small arteries on the SD image.
Motion artifacts are prevalent problems;however, the simple pixel-shifting methodused in x-ray DSA cannot be generalizeddirectly to MR DSA, in which data are ac-quired in complex k space. Motion effectscan be corrected for by detecting motiondisplacement and compensating for thephase shifts for each individual echo, a taskof complexity beyond the scope of this ar-ticle (22). The simple linearly filtered imageprovides reduction in motion artifacts andimprovement in vascular delineation, andit may serve as a simple and practical toolfor summarizing images. The linear filteringcan be regarded as motion gating in thatimages with substantial motion are dis-carded from summation. Currently, we areworking on motion detection algorithmsto fully automate motion gating. In sum-mary, postprocessing methods such as re-masking and filtering can be used to reducemotion effects and summarize vascularanatomy on a single high CNR image ob-tained by means of dynamic 2D MR DSA.
Cope C. Magnetic resonance imaging of an-giographically occult runoff vessels in pe-ripheral arterial occlusive disease. N EnglJ Med 1992; 326:1577–1581.
3. Korosec FR, Frayne R, Grist TM, Mistretta CA.Time-resolved contrast-enhanced 3D MR an-giography. Magn Reson Med 1996; 36:345–351.
4. Prince MR, Grist TM, Debatin JF. 3D contrastMR angiography. 2nd ed. New York, NY:Springer-Verlag, 1998.
5. Rofsky NM, Adelman MA. MR angiographyin the evaluation of atherosclerotic periph-eral vascular disease. Radiology 2000; 214:325–338.
6. Ho KY, Leiner T, de Haan MW, Kessels AG,Kitslaar PJ, van Engelshoven JM. Peripheralvascular tree stenoses: evaluation with mov-ing-bed infusion-tracking MR angiography.Radiology 1998; 206:683–692.
7. Wang Y, Lee HM, Khilnani NM, et al. Bolus-chase MR digital subtraction angiography inthe lower extremity. Radiology 1998; 207:263–269.
8. Wang Y, Lee HM, Avakian R, Winchester PA,Khilnani NM, Trost D. Timing algorithm forbolus chase MR digital subtraction angiogra-phy. Magn Reson Med 1998; 39:691–696.
9. Ho VB, Choyke PL, Foo TK, et al. Automatedbolus chase peripheral MR angiography: ini-tial practical experiences and future direc-tions of this work-in-progress. J Magn ResonImaging 1999; 10:376–388.
10. Meaney JF, Ridgway JP, Chakraverty S, et al.Stepping-table gadolinium-enhanced digitalsubtraction MR angiography of the aorta andlower extremity arteries: preliminary experi-ence. Radiology 1999; 211:59–67.
11. Ruehm SG, Hany TF, Pfammatter T, Schnei-der E, Ladd M, Debatin JF. Pelvic and lowerextremity arterial imaging: diagnostic perfor-mance of three-dimensional contrast-en-hanced MR angiography. AJR Am J Roentge-nol 2000; 174:1127–1136.
12. Earls JP, DeSena S, Bluemke DA. Gadolinium-enhanced three-dimensional MR angiogra-phy of the entire aorta and iliac arteries withdynamic manual table translation. Radiol-ogy 1998; 209:844–849.
13. Wang Y, Winchester PA, Khilnani NM, et al.Contrast enhanced peripheral MRA from ab-dominal aorta to pedal arteries: combineddynamic 2D and bolus chase 3D acquisi-tions. Invest Radiol 2001; 36:170–177.
14. Wang Y, Johnston DL, Breen JF, et al. Dy-namic MR digital subtraction angiographyusing contrast enhancement, fast data acqui-sition, and complex subtraction. Magn Re-son Med 1996; 36:551–556.
16. Winchester PA, Lee HM, Khilnani NM, et al.Comparison of two-dimensional MR digitalsubtraction angiography of the lower ex-tremity with x-ray angiography. J Vasc IntervRadiol 1998; 9:891–900.
17. Kruger RA, Riederer SJ. Digital subtractionangiography. Boston, Mass: Hall Medical,1984; 197–219.
18. Kruger RA. A method for time domain filter-ing using computerized fluoroscopy. MedPhys 1981; 8:466–470.
19. Wang Y, Weber DM, Korosec FR, et al. Gen-eralized matched filtering for time-resolvedMR angiography of pulsatile flow. Magn Re-son Med 1993; 30:600–608.
Figure 3. Coronal 2D MR DSA (10/2; flip an-gle, 60°) images obtained in a 24-year-old manwith a nonhealing ulcer. A, Arterial phase im-age in which the default mask was used. B,Peak arterial phase image (corresponding to A)in which an optimized mask was used. C, Fil-tered image with SD from the first mask to thelast arterial phase images. D, Linearly filteredimage in which manually selected masks (ninenonenhanced images) and arterial phases (fourarterial phases) were used. Both filtered imagesC and D) demonstrate substantial improve-ment in vascular CNR.
