- Email: [email protected]
Subtraction Computed Tomographic Angiography of Calcified Arteries
Subtraction Computed Tomographic Angiography of Calcified Arteries: Preliminary Phantom and Clinical Studies1 Peter J. Yim, PhD, John L. Nosher, MD, Anthony Burgos, MD, Ihab Haddadin, MD
Rationale and Objectives. The technique of subtraction computed tomographic angiography (sCTA) has been proposed for the evaluation of atherosclerotic disease to address limitations in CTA in highly calcified arteries. However, sCTA has not gained acceptance in clinical practice, in part, due to image artifacts caused by patient motion that occur between the acquisition of the two component images. The purpose of this study was to evaluate the effectiveness of computational image co-registration to obtain sCTA. Materials and Methods. The study was conducted using a semi-automated implementation of the mutual information (MI) registration algorithm. The results of sCTA were evaluated quantitatively in a phantom representing a calcified artery. Technical success of sCTA was evaluated in 14 calcified arterial segments in two patients. An observer study was carried out to determine interobserver agreement in the interpretation of sCTA. Qualitative observations were made between sCTA and CTA. Results. Computation time for performing the co-registration for each 2-cm calcification is less than 1 second. The necessary user interaction required minimal expertise. Measurements of the degree of stenosis in the calcified artery phantom agreed to within 8 ⫾ 4% of gold-standard measurements. Technical success was demonstrated in all calcifications. Strong interobserver agreement was obtained for the detection of hemodynamically significant stenoses ( ⫽ 0.86). Several apparent pitfalls in the interpretation of CTA in calcified arteries were noted that could potentially be obviated by sCTA. Conclusions. The study supports the use of a straight-forward implementation of the MI algorithm and provides preliminary evidence validating the use of sCTA in the setting of atherosclerotic disease of the lower extremities. Key Word. Computed tomographic angiography; image registration; mutual information; peripheral vascular disease ©
AUR, 2009
The accuracy of the computed tomographic angiographic (CTA) examination may be degraded by the presence of calcified plaque. Detection and quantification of such calcification may have value in the diagnosis and therapy planning of this disease (1–3); however, in relation to the primary criteria of disease severity (the degree of stenoAcad Radiol 2009; 16:257–265 1
From the Department of Radiology, UMDNJ-Robert Wood Johnson Medical School, Medical Education Building 404, New Brunswick, NJ 08903. Received March 19, 2008; revised June 18, 2008; accepted July 10, 2008. Address correspondence to: P.J.Y. e-mail: [email protected]
©
AUR, 2009 doi:10.1016/j.acra.2008.07.003
sis), such calcifications are simply a nuisance. Calcifications appear in CTA as small objects with a high computed tomographic density that can sometimes approach that of bone. In the cross-sectional view, the lumenal region of the calcified artery may appear distinctive as a region with intermediate intensity surrounded either partially or completely by a high-intensity region corresponding to the calcification. Thus, common radiologic practice for assessing calcified arteries in CTA is to review the cross-sectional images. However, even with this approach, recent studies suggest that detection of hemodynamically significant stenoses in the presence of calcified plaque can be inaccurate. For example, in the study of Willmann
257
YIM ET AL
et al (4), the presence of calcification tended to cause an overestimation of the degree of stenosis. In their study of 39 consecutive patients with 35 hemodynamically significant stenoses, overestimation of stenosis occurred in 26 vessel segments. In 20 of the cases in which the stenosis was overestimated, the primary cause of overestimation of stenosis was the presence of calcification. Ouwendink et al (5) found that wall calcifications in CTA often limited the diagnostic value of CTA and were a statistically significant predictor of when a patient would need additional imaging studies. In the study presented here, we elaborate on a promising new approach to computed tomographic angiography involving, in essence, digital subtraction of the non-contrast computed tomography from the computed tomography angiography. Our study addresses practical considerations that arise for implementation of this technique including, most importantly, the need to correct for patient motion that occurs between the acquisition of the noncontrast CT (ncCT) and the computed tomographic angiography. A computational image registration algorithm plays a large role in this respect. The study also involves robust validation of the methodology including a study using a realistic phantom as well as observations from a clinical study.
MATERIALS AND METHODS Patients This study was based on imaging from two patients who underwent a combined CTA and ncCT study for suspicion of arterial occlusive disease of the lower extremities. Analysis of the images was approved by our institutional review board. Acquisition of the combined CTA and ncCT was performed for clinical considerations in these patients. Calcified-artery Phantom Subtracted computed tomographic angiography (sCTA) was simulated using a vascular phantom. The phantom was constructed to include four relevant components of calcified arteries, the including blood pool, vessel wall, calcification, and extravascular tissue. Construction of the phantom was performed in-house. The vessel wall was formed from a polypropylene straw with a wall thickness of 0.05 mm. The phantom consisted of two stenosis models. The stenoses were concentric in cross-sectional shape and cross-sectional position. Calcification was formed
258
Academic Radiology, Vol 16, No 3, March 2009
with plasticine clay that, as applied to the phantom, has a maximum computed tomographic density of 1230 Hounsfield Units (HU) at 120 kVp, which is close to that of highly calcified plaque. The clay was applied in a thin layer to the outside of the vessel wall at the locations of the stenoses, with a maximal thickness of approximately 2 mm. The shape of the calcifications was irregular, and they were applied asymmetrically in relation to the axis of the stenosis. The area surrounding the vessel, corresponding to soft tissue was filled with water. The blood pool was modeled alternately with water and with a sugar-based solution that produces a computed tomographic density of 300 HU at 120 kVp at the center of the lumen. The water and the sugar-based solution represent noncontrast and contrast-filled conditions respectively. Flow connections were made to a syringe to allow for exchange of the two blood pool fluids during the imaging study. The phantom is shown schematically in Figure 1. Image Acquisition: Vascular Phantom All computed tomographic images of the phantom were acquired with a four-detector row Lightspeed Plus CT (GE Medical Systems, Waukesha, WI). Images were acquired in one session in which the technique of sCTA was simulated and in a second session in which gold standard images were obtained. In the sCTA session, the vascular phantom was imaged in two phases. In the first phase, the lumenal space was filled with water to simulate blood alone and in then second phase, the lumenal space was filled with a sugar solution. The phantom was shifted and turned slightly between the two phases of the image acquisition to simulate patient motion. The image acquisition in both phases was identical using the “runoff ” protocol. Parameters of the acquisition included a tube voltage of 120 kVp, a tube current of 150 mA, a reconstruction slice thickness of 1.25 mm, and a 35-cm reconstruction field-of-view. In the second session, gold standard images were acquired. In this session the clay, simulating calcified plaque, was removed from the phantom. Also, the lumenal space was filled with the sugar solution. Image acquisition was performed with the “routine head” protocol. Parameters of the acquisition included a tube voltage of 140 kVp, a tube current of 255 mA, a reconstruction slice thickness of 1.25 mm, and a 17.5 cm field-of-view. Image Acquisition: Clinical Cases Imaging was performed on a 64-detector-row Lightspeed VCT. The images were acquired with the standard-
Academic Radiology, Vol 16, No 3, March 2009
SUBTRACTION COMPUTED TOMOGRAPHIC ANGIOGRAPHY
Figure 1. Schematic of the calcified artery phantom. The phantom represents the relevant four elements of the imaging field: lumen (L), calcification (C), vessel wall (VW), and extravascular space (EVS). The phantom was constructed with two stenoses of varying stenosis length (SL) and stenosis diameter (SD).
of-care clinical protocol that included a tube voltage of 120 kVp and a tube current of 101 mA and were reconstructed at a slice thickness of 1.25 mm and an in-plane resolution of 0.75 mm. Images were acquired with a helical acquisition with a pitch of 1.375. Motion of the subject was not restricted during the computed tomographic examination, and the subject was not requested to remain motionless between acquisition of the ncCT and the CTA. Image Registration Some patient motion is expected to occur between the two phases of the subtraction CTA acquisition, even if the patient is instructed not to move. Thus, in general, a correction must be applied to one of the images to compensate for that motion. A semi-automated co-registration algorithm was used in this study. The algorithm includes manual registration, manual identification of calcifications and computational registration. Details of the registration process are given in the following. The first step in the registration process is manual registration of the noncontrast computed tomography and the CTA. Manual registration is expected to be relatively reliable because interscan motion will usually be small scale and primarily translational in nature. Thus, corresponding features in the two images can be recognized without extensive searching. Furthermore, the structures that are of most interest are the calcifications that have distinctive small-scale features that can be matched between the two images. Thus, the manual registration process for a given arterial territory consists of identifying a matching pair of points from the two images on a calcification. One of the images is then shifted both in- and out-of-plane to produce alignment of those matching points. In the proposed registration algorithm, the entire images are not registered to one another because the spatial transformation or motion correction between the two images may become complex for large volumes of tissue, including elastic deformations. Instead, the images are
registered to one another in segments with relatively limited volumes. Within such volumes, assumptions of the rigidity of the tissue are valid or nearly so. More specifically, these registration subvolumes are defined to encompass calcifications and ideally are positioned such that the calcifications are centered within the volumes. The accuracy of the subsequent computational registration will depend to some extent on whether there are an adequate number of points within the calcification within the subvolume and on whether those points are sufficiently centered within the subvolume. These subvolumes can be obtained in the following manner. First, the user places a series of points along a calcified segment in the artery with each point separated from the previous point by approximately 2 cm. This series of points defines the path of a calcified segment of an artery and also subdivides the artery in the axial direction. Each pair of consecutive points in this series can then be used to define a bounding box that surrounds the vessel and calcifications. First, a box is defined by considering the two points as its diametrically opposed corners. Then, that box is extended outwards by 1 cm in all directions to form the subvolume bounding box. Thus, the bounding box has a minimum possible dimension of 2 cm in any direction and a maximum dimension of 4 cm. Computational registration is then applied sequentially to each of the subvolumes within the respective bounding boxes. The process of constructing the subvolumes for registration is shown schematically in Figure 2. In our study, all interactive aspects of the registration process were performed in MIPAV (National Institutes of Health, Bethesda, MD) including identifying points for manual registration, identifying points along the calcified vessel axes, and cropping of the images. Corresponding subvolumes of the noncontrast computed tomography and the CTA were then co-registered using the mutual-information algorithm (6). The software algorithm was implemented in-house in “C” on the IRIS
259
YIM ET AL
Academic Radiology, Vol 16, No 3, March 2009
an average of one image-lumen/calcification point per bin in the joint histogram. Registration was based on a rigid model of motion with both translational and rotational components. Intraregistration resampling was performed using trilinear interpolation. In the registration process, the CTA served as the reference image. whereas the noncontrast computed tomogram served as the floating image. Optimization was performed using the gradient-descent method with translational increments of 0.1 mm in each direction and rotational increments of 0.2° around each axis. Based on the results of the mutual information registration, the optimal motion correction was applied to the entire noncontrast computed tomographic subvolume and then was arithmetically subtracted from the CTA subvolume. The resulting subtraction subvolume was then reinserted into its original position in the CTA after cropping a margin of 0.5 cm, where subtraction may not have been effective due to nonoverlap of the CTA sub-volume and the noncontrast computed tomographic subvolume.
Figure 2. Image regions included in co-registration of noncontrast computed tomography and the computed tomographic angiography. The regions are based on a series of two or more points along the calcified artery identified by the user. A box is then constructed, such as a pair of consecutive points (P1, P2), such that those points are located at the diametrically opposite corners. The box is then extended outwards by 0.5 and 1.0 cm to form image subvolumes in computed tomographic angiography (SVCTA) and in noncontrast computed tomogaphy (SVNCCT).
Explorer prototyping platform (Numerical Algorithms Group Ltd, Oxford, UK) on a Windows 2000 workstation. Preprocessing of the subvolumes included resampling to increase the resolution by a factor of two using a third-order b-spline function as implemented in MIPAV (National Institutes of Health). Within each CTA subvolume, a set of points was defined that were expected to be either lumen or calcification. These points were identified by thresholding of the CTA at a level of 100 HU and by manual contouring to exclude possible regions of bone in the subvolumes. Points within a 0.5-cm margin in the CTA were also excluded from this set to ensure complete overlap with the noncontrast CT for a wide range of translations and rotations. Calculations of mutual information were then based entirely on these lumen/calcification points. The histographic bin size for computing the mutual information cost function was set, such that there was
260
Image Analysis: Vascular Phantom sCTA of the phantom was rendered with the maximum intensity projection (MIP) at a series of views perpendicular to the native slice plane. The views spanned 180° at 30° increments. Image resampling for performing the rotations was done using trilinear interpolation. The orthonormal perspective was used for creating the MIPs. The series of MIPs was then viewed by researcher (P.J.Y.) using MIPAV (National Institutes of Health) to select a point of maximal stenosis in one of the views. Lines for forming image-intensity profiles were drawn at the point of maximal stenosis and at a normal segment of the model using the line tool. An edge enhancement filter, the gradient magnitude, was then applied to the MIP using MIPAV. The convolution kernel of the filter was set to a space constant of 1.0 ⫻ 1.0 pixels. Imageintensity profiles of the gradient-magnitude image at the stenotic and normal locations were then analyzed to determine the respective lumenal diameters. Specifically, the diameter was considered to be the distance between the two highest maxima in the image intensity profiles corresponding to the two opposing sides of the lumen. A similar process was used for measuring diameters in the gold standard images although profiles were obtained from the cross-sectional imaging. The effect of initialization error on the accuracy of the registration was also evaluated. The initialization position of the floating image was varied by displacements of up to 5.0
Academic Radiology, Vol 16, No 3, March 2009
SUBTRACTION COMPUTED TOMOGRAPHIC ANGIOGRAPHY
Figure 3. Subtraction computed tomographic angiography of calcified artery phantom. The images represent the results of simulation of the computed tomographic angiography (a,e), the noncontrast computed tomography (b,f), and the subtraction computed tomographic angiography (c,g). Reference images (d,h) were obtained from imaging of the phantom without calcification material present.
mm in each direction from the displacement obtained by registration using a best guess initialization. Observer Study: Clinical Cases sCTA of clinical cases was reviewed by two radiology residents (A.B., I.H.), for detection of hemodynamically significant stenoses. Such stenoses were defined as points with
greater than or equal to 50% narrowing that could be identified with at least moderate confidence. Rendering of the clinical sCTA was performed in a similar manner to that of the phantom study. Rendering sCTA was performed after removal of bone using manual contouring. Twelve calcified arterial segments were included in the analysis from the superficial femoral artery (n ⫽ 3), pop-
261
YIM ET AL
Academic Radiology, Vol 16, No 3, March 2009
Table 1 Results from the Phantom Study: Measurements from Subtracted Computed Tomograpic Angiography (sCTA) Compared with Gold Standard (GS) Measurements
Model #1 Model #2 Mean ⫾ SD
GS Stenosis (%)
sCTA Stenosis (%)
Absolute Error (%)
32 49 —
27 60 —
5 11 8⫾4
SD, standard deviation.
Figure 4. Effect of transformation error in the initialization of the image registration on the accuracy of the image transformation following registration. Only the translational component of the image transformation is represented in the error. The true image transformation was estimated as that obtained from a best guess initialization followed by mutual information registration.
liteal artery (n ⫽ 2), and popliteal trifurcation arteries (n ⫽ 7). The total length of calcified segments of arteries included in the study was approximately 28 cm. The image co-registration and subtraction was performed for fourteen 2-cm calcification subvolumes. RESULTS A high degree of suppression of the calcification component was achieved in the phantom study as is shown in Figure 3. Residual artifact from the calcification was seen but was significantly lower in intensity than the lumen and did not obstruct the view of the lumen in the MIP. Measurement of the degree of stenosis from the sCTA was within 8 ⫾ 4% for the two models (Table 1). The registration process was found to be insensitive to the initialization for error in the initialization position of up to 2.0 mm. Significant registration error was found to occur for larger initialization errors as shown in Figure 4. sCTA was considered to be technically successful in all 14 2-cm calcifications as judged (by P.J.Y.) to produce an arterial segment with a realistic appearance. An example of sCTA is shown in Figure 5. A high degree of interobserver agreement was also obtained for the detection of hemodynamically significant stenoses ( ⫽ 0.86). A breakdown of the outcome combinations is shown in Table 2.
262
Table 2 Breakdown of Outcomes from Interobserver Agreement Study for the Detection of Stenosis
Negative Positive Total
Negative
Positive
Total
6 1 7
0 5 5
6 6 12 segments
Negative, number of arterial segments in which stenoses was not detected; Positive, number of arterial segments in which stenosis was detected.
Practical aspects of the clinical use of sCTA include the degree of user interaction and computation time. The use of this implementation of sCTA was found to require modest user interaction that primarily relates to the identification of corresponding points in the ncCT and the CTA for performing manual registration. Identification of such points was not found to be problematic in the cases included in this study. The identification of one point within a given vascular territory was found to be adequate. Manual registration had to be repeated in one of the three vascular territories when obvious misregistration was seen in the sCTA. The computation time for performing sCTA was approximately 9 seconds for each 2-cm calcification, including 8 seconds for image interpolation for each 2-cm calcification and 1 second for mutual information (MI) image co-registration. Possible errors in the interpretation of CTA were noted that could potentially be resolved by using sCTA. These include: (1) the underestimation of the degree of stenosis due to the presence of calcified plaque that is isodense with the lumen, and (2) the false impression of vascular occlusion associated with a concentric calcification. These effects are shown in Figure 6.
Academic Radiology, Vol 16, No 3, March 2009
SUBTRACTION COMPUTED TOMOGRAPHIC ANGIOGRAPHY
Figure 5. Subtraction computed tomographic angiography in the territory of the superficial femoral artery. Subtraction computed tomographic angiography (c) obtained from computed tomographic angiography (a) and noncontrast computed tomography (b).
Figure 6. Types of calcified arterial segments in which computed tomographic angiography is likely to be misinterpreted. One such type is where the calcification is highly inhomogenous as shown in the computed tomographic angiogram (b) and noncontrast computed tomography (c) and whose position is indicated by the top line in (a). Another type is where the calcification concentrically surrounds the lumen as shown in the computed tomographic angiogram (d) and the noncontrast computed tomography (e) and whose position is indicated by the bottom line in (a). In the first type of arterial segment, the degree of stenosis may be underestimated in computed tomographic angiography. In the second type of arterial segment, total occlusion may be suspected based on reading of the computed tomographic angiogram.
DISCUSSION The present study strongly suggests the potentially accurate and robust application of sCTA for the suppression of calcification. In particular, the use of mutual informa-
tion registration appears to be justified in this context. This study builds on earlier reports of the use of sCTA by providing more robust validation as well as a more practical framework for clinical implementation. Highlights of our findings include the validation of the accuracy of
263
YIM ET AL
sCTA in a vascular phantom. Also, results from a clinical study shows that sCTA produces images that have a highly realistic appearance that can be interpreted in a reproducible manner. Also, possible limitations of conventional CTA in visualization of the lumen were noted where sCTA may be advantageous. Other important findings relate to the practicality of sCTA. Our study shows that this technique can be implemented in a manner that is fully compatible with clinical practice. In particular, the study proposes a semi-automated technique for performing sCTA that involves manual co-registration of the noncontrast computed tomogram and the CTA for each vascular territory and for the user to trace the path of each calcified arterial segment. These steps are not time-consuming and do not require significant expertise. Finally, the computational requirements of sCTA are relatively minor and thus the process can be carried out virtually in real time. The registration process has been found to be essentially insensitive to error in the manual initialization of the registration of up to 2.0 mm. The preliminary studies on clinical cases suggest that this degree of accuracy in the manual registration is feasible. There have been several previous studies addressing the prospects of sCTA. A major focus of all of these studies was on co-registration of the noncontrast computed tomography and CTA. Poletti et al (7) found a high rate of truepositive findings with sCTA (95.9%) in comparison with digital subtraction angiography. However, the technique of Poletti et al required that patient motion be minimized by placing restraints on the legs of the patient. Even so, diagnostic use of the sCTA in this study was not feasible in 20% of cases, presumably due to excessive patient motion. Registration of those images was performed in an interactive manner using in-house software. The software allowed for correction of translational motion. A single translation was applied to the entire image. The use of a computational algorithm for co-registration of the noncontrast computed tomography and the CTA offers the potential of improving the precision of the result over what can be obtained by manual methods. The appropriate approach to performing this registration is very much up for debate. The core of the registration process, the cost function, that represents the criteria for defining correct alignment, has been defined variously. Kwon et al (8) and Loeckx et al (9) have used the mutual information cost function that has been proposed in widespread applications both within and outside of medical imaging for image co-registration. This cost function has been found to be particularly useful for the co-registration
264
Academic Radiology, Vol 16, No 3, March 2009
of images obtained with different modalities, where the images have different appearances and may even have reversals in the relative contrasts between objects in the image. An example of this type of application is in the registration of positron emission tomography and computed tomography (10). In the case of sCTA, obviously the two images are acquired with the same modality, but, in the vicinity of the arteries, the images have a very distinctive appearance due to the absence of lumenal contrast in one and the presence in the other. Mutual information between two images is maximized, in general terms, when a given intensity in one image corresponds with a high probability to another, potentially different, intensity in the second image. Other criteria that have been considered for image registration in sCTA include the deterministic sign-change (DSC) criteria (11), a maximum cancellation (MC) cost function (12) and a modified least-squares (MLS) criteria (13). Of these, MC and MLS have been applied to sCTA for the purpose of suppression of calcification. It has not been shown as to whether these different registration criteria produce any substantive differences in the registration accuracy. In this study, the use of MI was chosen simply because its underlying principle and applicability are somewhat more general than that of the other techniques. Also, it was noted in qualitative preliminary testing that the results using MI were at least equivalent to those obtained by MC. Study Limitations Although the results of the present study are very promising, the study is preliminary in nature, involving only a phantom study and a limited number of clinical cases. Also, even in the clinical cases where the technique was evaluated, a clear standard-of-comparison, such as intra-arterial digital subtraction angiography was not available. Further, more extensive clinical studies of this technique will certainly be needed to clarify the its reliability and accuracy. Conclusions A very promising option for co-registration of the component images for forming sCTA has been developed and validated. The method has only moderate computational and user-interaction demands that are well within the usual constraints of clinical practice. The co-registration process was designed to be as simplistic as possible, including the use of a piecewise-rigid model of patient motion, gradient-descent optimization, and a generic ver-
Academic Radiology, Vol 16, No 3, March 2009
SUBTRACTION COMPUTED TOMOGRAPHIC ANGIOGRAPHY
sion of the MI cost function. This approach was seen to be effective. However, some artifacts in the sCTA, particularly in the phantom study, are likely due to residual misregistration and may be subject to further improvement. Also, manual interaction, particularly for initialization of the co-registration, plays an important role in obtaining sCTA. We believe this is probably acceptable in clinical practice but ideally, the user interaction should be further reduced or eliminated. REFERENCES 1. Rutten A, Isgum I, Prokop M. Coronary calcification: effect of small variation of scan starting position on Agatston, volume, and mass scores. Radiology 2008; 246:90 –98.. 2. Suzuki M, Ozaki Y, Komura S, Nakanishi A. Intracranial carotid calcification on CT images as an indicator of atheromatous plaque: analysis of high-resolution CTA images using a 64-multidetector scanner. Radiat Med 2007; 25:378 –385. 3. Maldonado TS. What are current preprocedure imaging requirements for carotid artery stenting and carotid endarterectomy: have magnetic resonance angiography and computed tomographic angiography made a difference? Semin Vasc Surg 2007; 20:205–215. 4. Willmann JK, Baumert B, Schertler T, et al. Aortoiliac and lower extremity arteries assessed with 16-detector row CT angiography: prospective comparison with digital subtraction angiography. Radiology 2005; 236:1083–1093.
5. Ouwendijk R, Kock MC, van Dijk LC, van Sambeek MR, Stijnen T, Hunink MG. Vessel wall calcifications at multi-detector row CT angiography in patients with peripheral arterial disease: effect on clinical utility and clinical predictors. Radiology 2006; 241:603– 608. 6. Pluim JP, Maintz JB, Viergever MA. Mutual-information-based registration of medical images: a survey. IEEE Trans Med Imaging 2003; 22: 986 –1004. 7. Poletti PA, Rosset A, Didier D, Bachmann P, Verdun FR, et al. Subtraction CT angiography of the lower limbs: a new technique for the evaluation of acute arterial occlusion. ASR Am J Roentgenol 2004; 183: 1445–1448. 8. Kwon SM, Kim YS, Kim T, Ra JB. Novel digital subtraction CT angiography based on 3D registration and refinement. SPIE Med Imaging 2004; 5370:1156 –1165. 9. Loeckx D, Drisis S, Maes F, Vandermeulen D, Marchal G, Suetens P. Removal of plaque and stent artifacts in subtraction CT angiography using nonrigid registration and a volume penalty. Conf Proc IEEE Eng Med Biol Soc 2005; 4:4294 – 4297. 10. Vogel WV, Schinagl DA, Van Dalen JA, Kaanders JH, Oyen WJ. Validated image fusion of dedicated PET and CT for external beam radiation therapy in the head and neck area. Q J Nucl Med Mol Imaging. 2008; 52:74 – 83. 11. Bani-Hashemi AR, Krishnan A, Samaddar S. Warped matching for digital subtraction of CT-angiography studies. Proceedings SPIE 1996; 2710:428 – 437. 12. Yim PJ, Nosher JL. Subtraction computed tomographic angiography: use of pre-contrast images for calcification removal. Proceedings SPIE 2006; 6143:805– 811. 13. van Straten M, Venema HW, Streekstra GJ, Reekers JA, den Heeten GJ, Grimbergen CA. Removal of arterial wall calcifications in CT angiography by local subtraction. Med Phys 2003; 30:761–770.
265
Rationale and Objectives. The technique of subtraction computed tomographic angiography (sCTA) has been proposed for the evaluation of atherosclerotic disease to address limitations in CTA in highly calcified arteries. However, sCTA has not gained acceptance in clinical practice, in part, due to image artifacts caused by patient motion that occur between the acquisition of the two component images. The purpose of this study was to evaluate the effectiveness of computational image co-registration to obtain sCTA. Materials and Methods. The study was conducted using a semi-automated implementation of the mutual information (MI) registration algorithm. The results of sCTA were evaluated quantitatively in a phantom representing a calcified artery. Technical success of sCTA was evaluated in 14 calcified arterial segments in two patients. An observer study was carried out to determine interobserver agreement in the interpretation of sCTA. Qualitative observations were made between sCTA and CTA. Results. Computation time for performing the co-registration for each 2-cm calcification is less than 1 second. The necessary user interaction required minimal expertise. Measurements of the degree of stenosis in the calcified artery phantom agreed to within 8 ⫾ 4% of gold-standard measurements. Technical success was demonstrated in all calcifications. Strong interobserver agreement was obtained for the detection of hemodynamically significant stenoses ( ⫽ 0.86). Several apparent pitfalls in the interpretation of CTA in calcified arteries were noted that could potentially be obviated by sCTA. Conclusions. The study supports the use of a straight-forward implementation of the MI algorithm and provides preliminary evidence validating the use of sCTA in the setting of atherosclerotic disease of the lower extremities. Key Word. Computed tomographic angiography; image registration; mutual information; peripheral vascular disease ©
AUR, 2009
The accuracy of the computed tomographic angiographic (CTA) examination may be degraded by the presence of calcified plaque. Detection and quantification of such calcification may have value in the diagnosis and therapy planning of this disease (1–3); however, in relation to the primary criteria of disease severity (the degree of stenoAcad Radiol 2009; 16:257–265 1
From the Department of Radiology, UMDNJ-Robert Wood Johnson Medical School, Medical Education Building 404, New Brunswick, NJ 08903. Received March 19, 2008; revised June 18, 2008; accepted July 10, 2008. Address correspondence to: P.J.Y. e-mail: [email protected]
©
AUR, 2009 doi:10.1016/j.acra.2008.07.003
sis), such calcifications are simply a nuisance. Calcifications appear in CTA as small objects with a high computed tomographic density that can sometimes approach that of bone. In the cross-sectional view, the lumenal region of the calcified artery may appear distinctive as a region with intermediate intensity surrounded either partially or completely by a high-intensity region corresponding to the calcification. Thus, common radiologic practice for assessing calcified arteries in CTA is to review the cross-sectional images. However, even with this approach, recent studies suggest that detection of hemodynamically significant stenoses in the presence of calcified plaque can be inaccurate. For example, in the study of Willmann
257
YIM ET AL
et al (4), the presence of calcification tended to cause an overestimation of the degree of stenosis. In their study of 39 consecutive patients with 35 hemodynamically significant stenoses, overestimation of stenosis occurred in 26 vessel segments. In 20 of the cases in which the stenosis was overestimated, the primary cause of overestimation of stenosis was the presence of calcification. Ouwendink et al (5) found that wall calcifications in CTA often limited the diagnostic value of CTA and were a statistically significant predictor of when a patient would need additional imaging studies. In the study presented here, we elaborate on a promising new approach to computed tomographic angiography involving, in essence, digital subtraction of the non-contrast computed tomography from the computed tomography angiography. Our study addresses practical considerations that arise for implementation of this technique including, most importantly, the need to correct for patient motion that occurs between the acquisition of the noncontrast CT (ncCT) and the computed tomographic angiography. A computational image registration algorithm plays a large role in this respect. The study also involves robust validation of the methodology including a study using a realistic phantom as well as observations from a clinical study.
MATERIALS AND METHODS Patients This study was based on imaging from two patients who underwent a combined CTA and ncCT study for suspicion of arterial occlusive disease of the lower extremities. Analysis of the images was approved by our institutional review board. Acquisition of the combined CTA and ncCT was performed for clinical considerations in these patients. Calcified-artery Phantom Subtracted computed tomographic angiography (sCTA) was simulated using a vascular phantom. The phantom was constructed to include four relevant components of calcified arteries, the including blood pool, vessel wall, calcification, and extravascular tissue. Construction of the phantom was performed in-house. The vessel wall was formed from a polypropylene straw with a wall thickness of 0.05 mm. The phantom consisted of two stenosis models. The stenoses were concentric in cross-sectional shape and cross-sectional position. Calcification was formed
258
Academic Radiology, Vol 16, No 3, March 2009
with plasticine clay that, as applied to the phantom, has a maximum computed tomographic density of 1230 Hounsfield Units (HU) at 120 kVp, which is close to that of highly calcified plaque. The clay was applied in a thin layer to the outside of the vessel wall at the locations of the stenoses, with a maximal thickness of approximately 2 mm. The shape of the calcifications was irregular, and they were applied asymmetrically in relation to the axis of the stenosis. The area surrounding the vessel, corresponding to soft tissue was filled with water. The blood pool was modeled alternately with water and with a sugar-based solution that produces a computed tomographic density of 300 HU at 120 kVp at the center of the lumen. The water and the sugar-based solution represent noncontrast and contrast-filled conditions respectively. Flow connections were made to a syringe to allow for exchange of the two blood pool fluids during the imaging study. The phantom is shown schematically in Figure 1. Image Acquisition: Vascular Phantom All computed tomographic images of the phantom were acquired with a four-detector row Lightspeed Plus CT (GE Medical Systems, Waukesha, WI). Images were acquired in one session in which the technique of sCTA was simulated and in a second session in which gold standard images were obtained. In the sCTA session, the vascular phantom was imaged in two phases. In the first phase, the lumenal space was filled with water to simulate blood alone and in then second phase, the lumenal space was filled with a sugar solution. The phantom was shifted and turned slightly between the two phases of the image acquisition to simulate patient motion. The image acquisition in both phases was identical using the “runoff ” protocol. Parameters of the acquisition included a tube voltage of 120 kVp, a tube current of 150 mA, a reconstruction slice thickness of 1.25 mm, and a 35-cm reconstruction field-of-view. In the second session, gold standard images were acquired. In this session the clay, simulating calcified plaque, was removed from the phantom. Also, the lumenal space was filled with the sugar solution. Image acquisition was performed with the “routine head” protocol. Parameters of the acquisition included a tube voltage of 140 kVp, a tube current of 255 mA, a reconstruction slice thickness of 1.25 mm, and a 17.5 cm field-of-view. Image Acquisition: Clinical Cases Imaging was performed on a 64-detector-row Lightspeed VCT. The images were acquired with the standard-
Academic Radiology, Vol 16, No 3, March 2009
SUBTRACTION COMPUTED TOMOGRAPHIC ANGIOGRAPHY
Figure 1. Schematic of the calcified artery phantom. The phantom represents the relevant four elements of the imaging field: lumen (L), calcification (C), vessel wall (VW), and extravascular space (EVS). The phantom was constructed with two stenoses of varying stenosis length (SL) and stenosis diameter (SD).
of-care clinical protocol that included a tube voltage of 120 kVp and a tube current of 101 mA and were reconstructed at a slice thickness of 1.25 mm and an in-plane resolution of 0.75 mm. Images were acquired with a helical acquisition with a pitch of 1.375. Motion of the subject was not restricted during the computed tomographic examination, and the subject was not requested to remain motionless between acquisition of the ncCT and the CTA. Image Registration Some patient motion is expected to occur between the two phases of the subtraction CTA acquisition, even if the patient is instructed not to move. Thus, in general, a correction must be applied to one of the images to compensate for that motion. A semi-automated co-registration algorithm was used in this study. The algorithm includes manual registration, manual identification of calcifications and computational registration. Details of the registration process are given in the following. The first step in the registration process is manual registration of the noncontrast computed tomography and the CTA. Manual registration is expected to be relatively reliable because interscan motion will usually be small scale and primarily translational in nature. Thus, corresponding features in the two images can be recognized without extensive searching. Furthermore, the structures that are of most interest are the calcifications that have distinctive small-scale features that can be matched between the two images. Thus, the manual registration process for a given arterial territory consists of identifying a matching pair of points from the two images on a calcification. One of the images is then shifted both in- and out-of-plane to produce alignment of those matching points. In the proposed registration algorithm, the entire images are not registered to one another because the spatial transformation or motion correction between the two images may become complex for large volumes of tissue, including elastic deformations. Instead, the images are
registered to one another in segments with relatively limited volumes. Within such volumes, assumptions of the rigidity of the tissue are valid or nearly so. More specifically, these registration subvolumes are defined to encompass calcifications and ideally are positioned such that the calcifications are centered within the volumes. The accuracy of the subsequent computational registration will depend to some extent on whether there are an adequate number of points within the calcification within the subvolume and on whether those points are sufficiently centered within the subvolume. These subvolumes can be obtained in the following manner. First, the user places a series of points along a calcified segment in the artery with each point separated from the previous point by approximately 2 cm. This series of points defines the path of a calcified segment of an artery and also subdivides the artery in the axial direction. Each pair of consecutive points in this series can then be used to define a bounding box that surrounds the vessel and calcifications. First, a box is defined by considering the two points as its diametrically opposed corners. Then, that box is extended outwards by 1 cm in all directions to form the subvolume bounding box. Thus, the bounding box has a minimum possible dimension of 2 cm in any direction and a maximum dimension of 4 cm. Computational registration is then applied sequentially to each of the subvolumes within the respective bounding boxes. The process of constructing the subvolumes for registration is shown schematically in Figure 2. In our study, all interactive aspects of the registration process were performed in MIPAV (National Institutes of Health, Bethesda, MD) including identifying points for manual registration, identifying points along the calcified vessel axes, and cropping of the images. Corresponding subvolumes of the noncontrast computed tomography and the CTA were then co-registered using the mutual-information algorithm (6). The software algorithm was implemented in-house in “C” on the IRIS
259
YIM ET AL
Academic Radiology, Vol 16, No 3, March 2009
an average of one image-lumen/calcification point per bin in the joint histogram. Registration was based on a rigid model of motion with both translational and rotational components. Intraregistration resampling was performed using trilinear interpolation. In the registration process, the CTA served as the reference image. whereas the noncontrast computed tomogram served as the floating image. Optimization was performed using the gradient-descent method with translational increments of 0.1 mm in each direction and rotational increments of 0.2° around each axis. Based on the results of the mutual information registration, the optimal motion correction was applied to the entire noncontrast computed tomographic subvolume and then was arithmetically subtracted from the CTA subvolume. The resulting subtraction subvolume was then reinserted into its original position in the CTA after cropping a margin of 0.5 cm, where subtraction may not have been effective due to nonoverlap of the CTA sub-volume and the noncontrast computed tomographic subvolume.
Figure 2. Image regions included in co-registration of noncontrast computed tomography and the computed tomographic angiography. The regions are based on a series of two or more points along the calcified artery identified by the user. A box is then constructed, such as a pair of consecutive points (P1, P2), such that those points are located at the diametrically opposite corners. The box is then extended outwards by 0.5 and 1.0 cm to form image subvolumes in computed tomographic angiography (SVCTA) and in noncontrast computed tomogaphy (SVNCCT).
Explorer prototyping platform (Numerical Algorithms Group Ltd, Oxford, UK) on a Windows 2000 workstation. Preprocessing of the subvolumes included resampling to increase the resolution by a factor of two using a third-order b-spline function as implemented in MIPAV (National Institutes of Health). Within each CTA subvolume, a set of points was defined that were expected to be either lumen or calcification. These points were identified by thresholding of the CTA at a level of 100 HU and by manual contouring to exclude possible regions of bone in the subvolumes. Points within a 0.5-cm margin in the CTA were also excluded from this set to ensure complete overlap with the noncontrast CT for a wide range of translations and rotations. Calculations of mutual information were then based entirely on these lumen/calcification points. The histographic bin size for computing the mutual information cost function was set, such that there was
260
Image Analysis: Vascular Phantom sCTA of the phantom was rendered with the maximum intensity projection (MIP) at a series of views perpendicular to the native slice plane. The views spanned 180° at 30° increments. Image resampling for performing the rotations was done using trilinear interpolation. The orthonormal perspective was used for creating the MIPs. The series of MIPs was then viewed by researcher (P.J.Y.) using MIPAV (National Institutes of Health) to select a point of maximal stenosis in one of the views. Lines for forming image-intensity profiles were drawn at the point of maximal stenosis and at a normal segment of the model using the line tool. An edge enhancement filter, the gradient magnitude, was then applied to the MIP using MIPAV. The convolution kernel of the filter was set to a space constant of 1.0 ⫻ 1.0 pixels. Imageintensity profiles of the gradient-magnitude image at the stenotic and normal locations were then analyzed to determine the respective lumenal diameters. Specifically, the diameter was considered to be the distance between the two highest maxima in the image intensity profiles corresponding to the two opposing sides of the lumen. A similar process was used for measuring diameters in the gold standard images although profiles were obtained from the cross-sectional imaging. The effect of initialization error on the accuracy of the registration was also evaluated. The initialization position of the floating image was varied by displacements of up to 5.0
Academic Radiology, Vol 16, No 3, March 2009
SUBTRACTION COMPUTED TOMOGRAPHIC ANGIOGRAPHY
Figure 3. Subtraction computed tomographic angiography of calcified artery phantom. The images represent the results of simulation of the computed tomographic angiography (a,e), the noncontrast computed tomography (b,f), and the subtraction computed tomographic angiography (c,g). Reference images (d,h) were obtained from imaging of the phantom without calcification material present.
mm in each direction from the displacement obtained by registration using a best guess initialization. Observer Study: Clinical Cases sCTA of clinical cases was reviewed by two radiology residents (A.B., I.H.), for detection of hemodynamically significant stenoses. Such stenoses were defined as points with
greater than or equal to 50% narrowing that could be identified with at least moderate confidence. Rendering of the clinical sCTA was performed in a similar manner to that of the phantom study. Rendering sCTA was performed after removal of bone using manual contouring. Twelve calcified arterial segments were included in the analysis from the superficial femoral artery (n ⫽ 3), pop-
261
YIM ET AL
Academic Radiology, Vol 16, No 3, March 2009
Table 1 Results from the Phantom Study: Measurements from Subtracted Computed Tomograpic Angiography (sCTA) Compared with Gold Standard (GS) Measurements
Model #1 Model #2 Mean ⫾ SD
GS Stenosis (%)
sCTA Stenosis (%)
Absolute Error (%)
32 49 —
27 60 —
5 11 8⫾4
SD, standard deviation.
Figure 4. Effect of transformation error in the initialization of the image registration on the accuracy of the image transformation following registration. Only the translational component of the image transformation is represented in the error. The true image transformation was estimated as that obtained from a best guess initialization followed by mutual information registration.
liteal artery (n ⫽ 2), and popliteal trifurcation arteries (n ⫽ 7). The total length of calcified segments of arteries included in the study was approximately 28 cm. The image co-registration and subtraction was performed for fourteen 2-cm calcification subvolumes. RESULTS A high degree of suppression of the calcification component was achieved in the phantom study as is shown in Figure 3. Residual artifact from the calcification was seen but was significantly lower in intensity than the lumen and did not obstruct the view of the lumen in the MIP. Measurement of the degree of stenosis from the sCTA was within 8 ⫾ 4% for the two models (Table 1). The registration process was found to be insensitive to the initialization for error in the initialization position of up to 2.0 mm. Significant registration error was found to occur for larger initialization errors as shown in Figure 4. sCTA was considered to be technically successful in all 14 2-cm calcifications as judged (by P.J.Y.) to produce an arterial segment with a realistic appearance. An example of sCTA is shown in Figure 5. A high degree of interobserver agreement was also obtained for the detection of hemodynamically significant stenoses ( ⫽ 0.86). A breakdown of the outcome combinations is shown in Table 2.
262
Table 2 Breakdown of Outcomes from Interobserver Agreement Study for the Detection of Stenosis
Negative Positive Total
Negative
Positive
Total
6 1 7
0 5 5
6 6 12 segments
Negative, number of arterial segments in which stenoses was not detected; Positive, number of arterial segments in which stenosis was detected.
Practical aspects of the clinical use of sCTA include the degree of user interaction and computation time. The use of this implementation of sCTA was found to require modest user interaction that primarily relates to the identification of corresponding points in the ncCT and the CTA for performing manual registration. Identification of such points was not found to be problematic in the cases included in this study. The identification of one point within a given vascular territory was found to be adequate. Manual registration had to be repeated in one of the three vascular territories when obvious misregistration was seen in the sCTA. The computation time for performing sCTA was approximately 9 seconds for each 2-cm calcification, including 8 seconds for image interpolation for each 2-cm calcification and 1 second for mutual information (MI) image co-registration. Possible errors in the interpretation of CTA were noted that could potentially be resolved by using sCTA. These include: (1) the underestimation of the degree of stenosis due to the presence of calcified plaque that is isodense with the lumen, and (2) the false impression of vascular occlusion associated with a concentric calcification. These effects are shown in Figure 6.
Academic Radiology, Vol 16, No 3, March 2009
SUBTRACTION COMPUTED TOMOGRAPHIC ANGIOGRAPHY
Figure 5. Subtraction computed tomographic angiography in the territory of the superficial femoral artery. Subtraction computed tomographic angiography (c) obtained from computed tomographic angiography (a) and noncontrast computed tomography (b).
Figure 6. Types of calcified arterial segments in which computed tomographic angiography is likely to be misinterpreted. One such type is where the calcification is highly inhomogenous as shown in the computed tomographic angiogram (b) and noncontrast computed tomography (c) and whose position is indicated by the top line in (a). Another type is where the calcification concentrically surrounds the lumen as shown in the computed tomographic angiogram (d) and the noncontrast computed tomography (e) and whose position is indicated by the bottom line in (a). In the first type of arterial segment, the degree of stenosis may be underestimated in computed tomographic angiography. In the second type of arterial segment, total occlusion may be suspected based on reading of the computed tomographic angiogram.
DISCUSSION The present study strongly suggests the potentially accurate and robust application of sCTA for the suppression of calcification. In particular, the use of mutual informa-
tion registration appears to be justified in this context. This study builds on earlier reports of the use of sCTA by providing more robust validation as well as a more practical framework for clinical implementation. Highlights of our findings include the validation of the accuracy of
263
YIM ET AL
sCTA in a vascular phantom. Also, results from a clinical study shows that sCTA produces images that have a highly realistic appearance that can be interpreted in a reproducible manner. Also, possible limitations of conventional CTA in visualization of the lumen were noted where sCTA may be advantageous. Other important findings relate to the practicality of sCTA. Our study shows that this technique can be implemented in a manner that is fully compatible with clinical practice. In particular, the study proposes a semi-automated technique for performing sCTA that involves manual co-registration of the noncontrast computed tomogram and the CTA for each vascular territory and for the user to trace the path of each calcified arterial segment. These steps are not time-consuming and do not require significant expertise. Finally, the computational requirements of sCTA are relatively minor and thus the process can be carried out virtually in real time. The registration process has been found to be essentially insensitive to error in the manual initialization of the registration of up to 2.0 mm. The preliminary studies on clinical cases suggest that this degree of accuracy in the manual registration is feasible. There have been several previous studies addressing the prospects of sCTA. A major focus of all of these studies was on co-registration of the noncontrast computed tomography and CTA. Poletti et al (7) found a high rate of truepositive findings with sCTA (95.9%) in comparison with digital subtraction angiography. However, the technique of Poletti et al required that patient motion be minimized by placing restraints on the legs of the patient. Even so, diagnostic use of the sCTA in this study was not feasible in 20% of cases, presumably due to excessive patient motion. Registration of those images was performed in an interactive manner using in-house software. The software allowed for correction of translational motion. A single translation was applied to the entire image. The use of a computational algorithm for co-registration of the noncontrast computed tomography and the CTA offers the potential of improving the precision of the result over what can be obtained by manual methods. The appropriate approach to performing this registration is very much up for debate. The core of the registration process, the cost function, that represents the criteria for defining correct alignment, has been defined variously. Kwon et al (8) and Loeckx et al (9) have used the mutual information cost function that has been proposed in widespread applications both within and outside of medical imaging for image co-registration. This cost function has been found to be particularly useful for the co-registration
264
Academic Radiology, Vol 16, No 3, March 2009
of images obtained with different modalities, where the images have different appearances and may even have reversals in the relative contrasts between objects in the image. An example of this type of application is in the registration of positron emission tomography and computed tomography (10). In the case of sCTA, obviously the two images are acquired with the same modality, but, in the vicinity of the arteries, the images have a very distinctive appearance due to the absence of lumenal contrast in one and the presence in the other. Mutual information between two images is maximized, in general terms, when a given intensity in one image corresponds with a high probability to another, potentially different, intensity in the second image. Other criteria that have been considered for image registration in sCTA include the deterministic sign-change (DSC) criteria (11), a maximum cancellation (MC) cost function (12) and a modified least-squares (MLS) criteria (13). Of these, MC and MLS have been applied to sCTA for the purpose of suppression of calcification. It has not been shown as to whether these different registration criteria produce any substantive differences in the registration accuracy. In this study, the use of MI was chosen simply because its underlying principle and applicability are somewhat more general than that of the other techniques. Also, it was noted in qualitative preliminary testing that the results using MI were at least equivalent to those obtained by MC. Study Limitations Although the results of the present study are very promising, the study is preliminary in nature, involving only a phantom study and a limited number of clinical cases. Also, even in the clinical cases where the technique was evaluated, a clear standard-of-comparison, such as intra-arterial digital subtraction angiography was not available. Further, more extensive clinical studies of this technique will certainly be needed to clarify the its reliability and accuracy. Conclusions A very promising option for co-registration of the component images for forming sCTA has been developed and validated. The method has only moderate computational and user-interaction demands that are well within the usual constraints of clinical practice. The co-registration process was designed to be as simplistic as possible, including the use of a piecewise-rigid model of patient motion, gradient-descent optimization, and a generic ver-
Academic Radiology, Vol 16, No 3, March 2009
SUBTRACTION COMPUTED TOMOGRAPHIC ANGIOGRAPHY
sion of the MI cost function. This approach was seen to be effective. However, some artifacts in the sCTA, particularly in the phantom study, are likely due to residual misregistration and may be subject to further improvement. Also, manual interaction, particularly for initialization of the co-registration, plays an important role in obtaining sCTA. We believe this is probably acceptable in clinical practice but ideally, the user interaction should be further reduced or eliminated. REFERENCES 1. Rutten A, Isgum I, Prokop M. Coronary calcification: effect of small variation of scan starting position on Agatston, volume, and mass scores. Radiology 2008; 246:90 –98.. 2. Suzuki M, Ozaki Y, Komura S, Nakanishi A. Intracranial carotid calcification on CT images as an indicator of atheromatous plaque: analysis of high-resolution CTA images using a 64-multidetector scanner. Radiat Med 2007; 25:378 –385. 3. Maldonado TS. What are current preprocedure imaging requirements for carotid artery stenting and carotid endarterectomy: have magnetic resonance angiography and computed tomographic angiography made a difference? Semin Vasc Surg 2007; 20:205–215. 4. Willmann JK, Baumert B, Schertler T, et al. Aortoiliac and lower extremity arteries assessed with 16-detector row CT angiography: prospective comparison with digital subtraction angiography. Radiology 2005; 236:1083–1093.
5. Ouwendijk R, Kock MC, van Dijk LC, van Sambeek MR, Stijnen T, Hunink MG. Vessel wall calcifications at multi-detector row CT angiography in patients with peripheral arterial disease: effect on clinical utility and clinical predictors. Radiology 2006; 241:603– 608. 6. Pluim JP, Maintz JB, Viergever MA. Mutual-information-based registration of medical images: a survey. IEEE Trans Med Imaging 2003; 22: 986 –1004. 7. Poletti PA, Rosset A, Didier D, Bachmann P, Verdun FR, et al. Subtraction CT angiography of the lower limbs: a new technique for the evaluation of acute arterial occlusion. ASR Am J Roentgenol 2004; 183: 1445–1448. 8. Kwon SM, Kim YS, Kim T, Ra JB. Novel digital subtraction CT angiography based on 3D registration and refinement. SPIE Med Imaging 2004; 5370:1156 –1165. 9. Loeckx D, Drisis S, Maes F, Vandermeulen D, Marchal G, Suetens P. Removal of plaque and stent artifacts in subtraction CT angiography using nonrigid registration and a volume penalty. Conf Proc IEEE Eng Med Biol Soc 2005; 4:4294 – 4297. 10. Vogel WV, Schinagl DA, Van Dalen JA, Kaanders JH, Oyen WJ. Validated image fusion of dedicated PET and CT for external beam radiation therapy in the head and neck area. Q J Nucl Med Mol Imaging. 2008; 52:74 – 83. 11. Bani-Hashemi AR, Krishnan A, Samaddar S. Warped matching for digital subtraction of CT-angiography studies. Proceedings SPIE 1996; 2710:428 – 437. 12. Yim PJ, Nosher JL. Subtraction computed tomographic angiography: use of pre-contrast images for calcification removal. Proceedings SPIE 2006; 6143:805– 811. 13. van Straten M, Venema HW, Streekstra GJ, Reekers JA, den Heeten GJ, Grimbergen CA. Removal of arterial wall calcifications in CT angiography by local subtraction. Med Phys 2003; 30:761–770.
265