British Journal of Radiology (2004) 77, 123-128
© 2004 British Institute of Radiology
doi: 10.1259/bjr/27339681
Assessment of a technique for 2D–3D registration of cerebral intra-arterial angiography
J V Byrne, FRCR
1
C Colominas, MSc
1
J Hipwell, PhD
2
T Cox, FRCR
3
J A Noble, DPhil
4
G P Penney, PhD
2 and
D J Hawkes, PhD
2
1 Department of Neuroradiology, Nuffield Department of Surgery, University of Oxford, Radcliffe Infirmary, Woodstock Road, Oxford OX2 6HE, 2 Division of Imaging Sciences, Kings College London, Thomas Guy House, Guys Hospital, London SE1 9RT, 3 National Hospital for Neurology & Neurosurgery, Queen Square, London WC1 3BG and 4 Department of Engineering Sciences, University of Oxford, 17 Parks Road, Oxford OX1 3PJ, UK
 |
Abstract
|
|---|
This study assesses the ability of a computer algorithm to perform automated 2D–3D registrations of digitally subtracted cerebral angiograms. The technique was tested on clinical studies of five patients with intracranial aneurysms. The automated procedure was compared against a gold standard manual registration, and achieved a mean registration accuracy of 1.3 mm (SD 0.6 mm). Two registration strategies were tested using coarse (128 x 128 pixel) or fine (256 x 256 pixel) images. The mean registration errors proved similar but registration of the lower resolution images was 3 times quicker (mean registration times 33 s, SD 13 s for low and 150 s SD 48 s for high resolution images). The automated techniques were considerably faster than manual registrations but achieved similar accuracy. The technique has several potential uses but is particularly applicable to endovascular treatment techniques.
|
Introduction
|
|---|
Computer reconstruction techniques are now routinely used to display medical images in three-dimensions (3D). Rotational angiography systems are in clinical use to collect and display subtraction X-ray angiography in 3D (3D-DSA). This technology is particularly relevant to interventional neuroradiology (INR) because endovascular treatments (EVT) are highly dependent on accurate pre-operative and perioperative displays of vessel morphology [1–3].
There are a variety of cerebrovascular conditions that are now treated by interventional neuroradiologists. These include intracranial aneurysms and arteriovenous malformations. Rotational 3D-DSA has proved most useful in the diagnosis and pre-embolisation evaluation of intracranial aneurysms [4–6]. Embolisation has recently been shown to be safer than neurosurgical clipping after subarachnoid haemorrhage (SAH) due to aneurysm rupture [7]. It requires complex manipulations in space to place thrombogenic coils in the aneurysm sac; the process is currently guided by 2D X-ray views. Pre-embolisation planning is aided by 3D displays [8], but 3D-DSA is too expensive in terms of radiation exposure and acquisition time for repeated perioperative imaging. We have developed 2D–3D registration software [9, 10] so that 3D-DSA acquired pre-embolisation can be rapidly updated with 2D-DSA views to improve the operator's control during INR procedures.
Automated 2D–3D DSA registration has other potential applications in INR. For instance, prior to aneurysm treatments, the operator selects so called "working projections" to monitor the patency of adjacent arteries during the introduction of embolisation devices. These optimum views are crucial since incorrect device placement may impair distal cerebral blood flow and cause tissue ischaemia or infarction. Frequently there are only a limited number of positions which provide this information, so the ability to reposition 2D views acquired on different occasions to a 3D model would aid comparisons and be particularly useful for follow-up imaging. Furthermore it allows confirmation of aneurysm dimensions from the more accurate 3D-DSA and potentially rapid registration could be used to stabilize real-time road map or fluoroscopy displays. This study was performed to test the accuracy of the technique against a manually registered gold standard on clinical 2D and 3D data.
|
Method
|
|---|
Image acquisition
Rotational DSA images were acquired on an AXIOM Artis FA/BA X-ray unit (Siemens, Forcheim, Germany) using 40 degrees s-1 rotation for 200 degrees (Dynavision, Siemens) and acquiring projections at 30 frames s-1 in 2.5 degree steps. Intra-arterial trans-catheter injections of radiographic contrast medium (Omnipaque 300, Nicomed Ltd, Oslo, Norway) were made at 2.5 ml s-1 into the cervical cranial arteries by mechanical pump. The X-ray parameters were: 70 kV, 155 mA, pulse width 10.0 ms and 1.20 µGy pulse-1. A total of 150 mask and angiography images in 10242 matrix were acquired and exported to a Leonardo workstation (Siemens) for reconstruction and 3D display.
Images from angiograms acquired during the INR of five patients were studied. Pre-treatment patient consent for the use of anonymised images for research purposes was obtained. One 3D-DSA and two 2D-DSA sequences (a and b) were acquired in each patient. Bi-plane 2D-DSA images were obtained on the same X-ray imager in frontal and lateral projections at 3 frames s-1 in 10242 matrix, after hand injections of radiographic contrast medium into the cervical cranial arteries. The X-ray parameters were 77 kV, 135 mA, pulse width 160 ms. Images were selected from the arterial phase of the two sequences during contrast media filling of cerebral arteries. These arterial phase frames were summated into a maximum arterial opacity image (2D-MAX) which was used for the subsequent image registration.
Registration algorithm
We have previously described a technique to register 3D images acquired by magnetic resonance angiography (MRA) with 2D-DSA of cerebral vessel images [10]. In this study we use a similar approach, but apply the algorithm to the registration of 3D-DSA and 2D-DSA images. We obtain a segmentation of the vasculature in the 3D-DSA image by applying an intensity threshold. This threshold is specified by the user but varies little between patients. Briefly the registration algorithm operates as follows.
The goal of registration is the calculation of the position (tx,ty,tz) and orientation (
x,y,z), of the 3D-DSA image with respect to the 2D-MAX image. In addition four intrinsic parameters (u0, v0, k1, k2), which encapsulate the perspective projection associated with the 2D-DSA acquisition, are required. The parameters (u0, v0) specify the x and y coordinates in mm of the normal from the imaging plane to the X-ray source and hence the degree to which the X-ray image is "skewed" by any misalignment of the X-ray intensifier. The dimensionless parameters (k1, k2) specify the magnification in x and y generated by the projection and hence quantify the amount of perspective distortion present [12]. These perspective parameters can be obtained by acquiring images of a suitable calibration object such as a cube. Alternatively they may be estimated using the source to image distance (SID) and the intensifier size/field of view (FOV) obtained from the X-ray set and assuming that the centre of the image intensifier is exactly perpendicular to the X-ray source. We have used the latter approach in this paper so that: 
|
We have estimated the root mean square (RMS) reprojection error introduced by this estimation of the perspective parameters to be typically less than 0.6 mm (see section: "Accuracy measurement" below for a description of the reprojection error).
The registration algorithm iteratively estimates the values of the extrinsic parameters by repeatedly repositioning the 3D data set and computing its projection onto the X-ray intensifier plane. At each iteration a total of 12 projection images, or digitally reconstructed radiographs (DRRs), are generated by perturbing each of the six extrinsic parameters by a small amount in positive and negative directions. Each of these DRRs is then compared with the 2D-MAX image by calculating the value of the gradient difference [13] similarity measure. Those parameters for which the value of the similarity measure increases in a particular direction are then adjusted accordingly prior to the next iteration. This procedure is repeated until a maximal value of the similarity measure is found.
To reduce processing time and improve the robustness of the algorithm, a multiresolution strategy is adopted such that the registration is initially performed with the DRR and 2D-MAX image subsampled to 128 x 128 pixels (a "coarse" registration). Once the similarity measure has been optimized at this resolution the dimensions of both images are doubled and the optimization repeated with the images at twice the resolution ("fine" registration).
Starting position
As described above, the registration algorithm operates by projecting the 3D-DSA image onto the 2D-MAX image and estimating their similarity. To ensure convergence of the search for the registration position however, it is necessary to provide the algorithm with an initial starting position 
in which the DRR and 2D-MAX images are at least approximately superimposed. For instance in the registration of MRA and DSA described in [10], we found that a greater than 85% success rate is obtained if the registration starting position is within ±8° and ±50 mm out-of-plane translation of the correct registration position (the in-plane translation is assumed to have been eliminated by a simple manual point-picking procedure).
In practice knowledge of the X-ray unit gantry angles at which the 2D-DSA sequence was acquired enables the approximate orientation of the 3D-DSA image to be estimated (subject to the degree of patient movement between the two acquisitions). On the basis of the gantry angles recorded when the 2D-DSA runs were acquired, 2D-DSA orientation estimates (for the primary, secondary and tertiary/in-plane angles) were within ±4° of the gold standard position for 7 of the 10 patient/runs and within ±8° for the remaining 3 runs.
Having obtained the approximate orientation of the 3D-DSA image, 
, it is a simple task for the user to set its position, 
, using an interactive graphical tool. This tool displays a variably transparent surface rendering of the vasculature (the 3D model), segmented from the 3D-DSA image. The 3D model is placed between the viewer, located at the X-ray source, and the 2D-MAX image positioned at the image intensifier. The perspective of this display is determined by the previously calculated intrinsic parameters, (u0,v0,k1,k2). Movement of the computer mouse can then be used to rapidly bring the 3D model into approximate alignment with the 2D-MAX image.
The user also chooses a point located on the surface of the 3D model, close to the lesion or vessel of interest. This point specifies the centre of rotation to be used by the registration algorithm and also defines the centre of an 80 mm radius, spherical volume of interest (VOI), which was subsequently used to estimate the accuracy of each registration. An experienced user can set an approximate starting position and specify a suitable centre of rotation in a matter of seconds.
Setting a "gold" standard
To estimate the accuracy of the registration algorithm, we manually generated gold standard registrations, 
, for each of the 10 2D-DSA sequences using the graphical tool described above. However, in contrast to the approximate positioning used to generate a starting position for the registration algorithm, in this case the orientation and position of the 3D model was adjusted to bring it into precise alignment with the 2D-MAX image. To aid this task the sensitivity of the position of the model and orientation to movement of the mouse was increased to allow progressively finer adjustment of the six extrinsic parameters.
All the gold standard registrations were performed by a single investigator (CC), after initial practice and without any time constraints. The individual registrations took at least 20 min.
Accuracy measurement
To determine the quality of a given automated registration using the registration algorithm, comparison was made to the gold standard. The difference between the two registrations was quantified as the RMS reprojection error [11], DRMS, and was calculated over a regular 3D lattice of points (20 mm spacing) within the 80 mm radius spherical. The calculation of DRMS is illustrated in Figure 1. Each point Pi(i=1,K,Nreproj) within the VOI is projected onto the imaging plane using the gold standard registration 
, to create a set of corresponding 2D points ui. These 2D points are then reprojected using a given registration T(x,y,z,tx,ty,tz), to produce lines Li. The reprojection distance, Di, is then equal to the minimum distance from this line to the original point Pi. For a perfect registration the line Li will pass directly back through Pi and the reprojection distance, Di, will be equal to zero. The RMS target reprojection error, DRMS, is then given by:
This parameter encapsulates the mean misalignment of points within the 3D volume of interest when projected onto the 2D-MAX image. For example a 1 mm lateral misalignment of the patient, (ty,tz), with respect to the gold standard, in an anteroposterior view, will result in a value of DRMS of 1 mm.
View larger version (14K):
[in this window]
[in a new window]
|
Figure 1. Reprojection distance modified from [11]. Each point Pi(i=1,...,Nreproj) within the volume of interest is projected onto the imaging plane using the gold standard registration matrix G, to create corresponding 2D points ui. These 2D points are then reprojected using a given transformation matrix T to produce lines Li. The reprojection distance, Di, is then equal to the minimum distance from this line to the original point Pi. For a perfect registration the line Li will pass directly back through Pi and the reprojection error distance, Di, will be equal to zero. The root mean square (RMS) target reprojection error, 
|
|
Precision of the gold standard
We used the RMS reprojection error, DRMS, to estimate the interobserver and intraobserver reproducibility of the gold standard.
Four observers repeated the gold standard alignment procedure for each 2D-MAX image, i.e. P1a, P1b, P2a,...,P5b. In each case the manual registration parameters, 
, were assigned to T and the value of DRMS calculated with respect to the gold standard. The interobserver reproducibility was expressed as the mean and standard deviation for each set of four observers' values in Table 1a.
View this table:
[in this window]
[in a new window]
|
Table 1. (a) Gold standard reproducibility. Root mean square reprojection error in mm (mean and standard deviation (SD)) for the registration of two data sets from five patients performed by four observers (interobserver values). (P1a=patient1, run a; P1b=patient1, run b; P2a=patient2, run a...etc; N=number of observations)
|
|
In addition the creator of the gold standard (CC) repeated this procedure a further six times and the RMS reprojection error was calculated for each of these manual registrations with respect to the gold standard. The mean and standard deviation for each set of six registrations—the intraobserver reproducibility—is given in Table 1b.
|
Results
|
|---|
The results of automated registrations of the 10 2D-MAX images against the 3D-DSA data are given in Tables 2 and 3. Each registration was performed a total of four times to enable the sensitivity of the algorithm to variations in the manual start position to be determined.
View this table:
[in this window]
[in a new window]
|
Table 2. Root mean square reprojection errors in mm (mean and standard deviation (SD)) for registration of the 10 data sets. Each registration was performed a total of four times. The reprojection error is given for the initial manual start position, after completion of the coarse registration and also for the fine registration. (P1a=patient1, run a; P1b=patient1, run b; P2a=patient2, run a...etc;)
|
|
View this table:
[in this window]
[in a new window]
|
Table 3. Execution times in seconds (mean and standard deviation (SD)) for registration of the 10 data sets, repeated four times. (P1a=patient1, run a; P1b=patient1, run b; P2a=patient2, run a...etc;)
|
|
Table 2
compares the RMS reprojection error for the coarse and fine registrations with the initial error value at the manually specified starting position. For seven out of the 10 2D-MAX images this error was reduced from between 2 mm and 4.5 mm to less than 2 mm. For two of the remaining three cases (P1a and P5a) the reprojection error of the start position was already less than 2 mm and for the final case (P2b) the mean RMS reprojection error was reduced from 3.8 mm to around 2.7 mm. There was very little variation between the results obtained for the coarse and fine registrations, with the difference between the mean RMS reprojection errors being less than or equal to 0.3 mm for all 10 cases. Registration of data showing an anterior communicating artery aneurysm is illustrated in Figure 2.
View larger version (143K):
[in this window]
[in a new window]
|
Figure 2. Example registration for patient 2, run A. These angiograms show a large anterior communicating artery aneurysm in an oblique frontal view after injection of the right internal carotid artery. The root mean square reprojection error in the start position (C) was 8.2 mm and after automated registration (D) 0.9 mm.
A: Maximum opacity arterial display subtraction angiography, "2D-MAX", image.
B: Gold standard registration performed manually in about 20 min.
C: Starting position for the registration.
D: Computed registration position achieved in 3 min.
|
|
The difference between the coarse and fine registration strategies is more apparent in the executions times listed in Table 3. These figures indicate that the coarse registration typically takes from 30 s to 60 s whilst the fine registration is around three times slower. These figures were obtained for a 1.7 GHz Intel Foster processor with 1 Gbyte of RAM.
|
Discussion
|
|---|
We have previously demonstrated the ability of this algorithm to produce accurate registrations of cerebral angiograms [10]. A property of intensity based algorithms such as ours is the need to provide a starting position in which the 3D and 2D images are at least approximately superimposed. In this paper we have demonstrated that such a starting position can be easily achieved using knowledge of the imager gantry position during acquisitions and with the minimum of interaction from the user.
Our results show that there is very little difference in the accuracy of registrations performed at the coarse resolution (128 by 128 pixels) when compared with registrations performed at the fine resolution (256 by 256 pixels). The overall accuracy with respect to the gold standard of both strategies (1.3 mm SD 0.6) is comparable with the intraobserver reproducibility (1.3 mm SD 0.5) and slightly less than the interobserver reproducibility (1.6 mm SD 0.8). All these errors have been calculated over an 80 mm radius spherical volume of interest within the 3D-DSA image. The overall reprojection errors reported here (1.3 mm) are similar to those reported in our previous study to register MRA and DSA (less than 1.3 mm for clinical images).
We have shown therefore that the algorithm can produce a registration to an accuracy that is comparable with our gold standard manual registration but in a fraction of the time—typically less than a minute for the former compared with upwards of 20 min for the manual registration.
The value of 3D-DSA has been described in previous studies [1, 3–5] and its utility confirmed by the number of imagers being installed and currently in use clinically. Acquisition of 3D volume data is slower and involves a greater radiation patient dose than 2D views, factors which limit its periprocedural use. 2D–3D image registration algorithms will allow aneurysm dimensions to be objectively and reproducibly measured using the 3D image in which their position and correct dimensions have been determined. This avoids the inaccurate procedure of introducing calibration objects into the 2D image and takes advantage of the greater visibility of aneurysms in the 3D-DSA image. Although fiducials can in principle be used to obtain a 2D–3D registration, external placement can be inaccurate (due to skin movement), invasive (if bone mounted) or unfeasible in practice due to the limited field of view of the 3D-DSA (fiducials should be visible in all frames of the rotation). Clearly internal placement of fiducials is highly invasive. An image driven algorithm such as presented here, removes the need for fiducial markers and can be used during procedures to aid accurate repositioning of 2D to 3D views, refreshment of working projections and compensation for small changes in patient position. Though it has not been shown that real-time 3D displays will improve the results, it can be assumed that increasing the operator's knowledge of the local anatomy of a lesion will increase their confidence when undertaking INR procedures.
Received for publication April 11, 2003.
Revision received June 16, 2003.
Accepted for publication September 3, 2003.
|
References
|
|---|
- Schumacher M, Kutluk K, Ott D. Digital rotational angiography in neuroradiology. AJNR Am J Neuroradiol 1989;10:644–9.[Medline]
- Heautot JF, Chabert E, Gandon Y, Croci S, Romeas R, Campagnolo R, et al. Analysis of cerebrovascular disease by a new 3-dimensional computerized X-ray angiography system. Neuroradiology 1998;40:203–9.[CrossRef][Medline]
- Anxionnat R, Bracard S, Macho J, Da Costa E, Vaillant R, Launay L, et al. 3D angiography: clinical interest—first application in interventional neuroradiology. J Neuroradiol 1998;25:251–62.[Medline]
- Hoff DJ, Wallace MC, terBrugge KG, Gentili F. Rotational angiography assessment of cerebral aneurysms. AJNR Am J Neuroradiol 1994;15:1945–8.[Abstract]
- Tu RK, Cohen WA, Maravilla KR, Bush WH, Patel NH, Eskridge J, et al. Digital subtraction rotational angiography for aneurysms of the intracranial anterior circulation: injection method and optimization. AJNR Am J Neuroradiol 1996;17:1127–36.[Abstract]
- Tanoue S, Kiyosue H, Kenai H, Nakamura T, Yamashita M, Mori H. Three-dimensional reconstructed images after rotational angiography in the evaluation of intra-cranial aneurysms: surgical correlation. Neurosurgery 2000;47:866–71.[CrossRef][Medline]
- International subarachnoid aneurysm trial (ISAT) collaborative group. International subarachnoid aneurysm trial (ISAT) of neurosurgical clipping versus endovascular coiling in 2143 patients with ruptured intracranial aneurysms: a randomised trial. Lancet 2002;360:1267–74.[CrossRef][Medline]
- Anxionnat R, Bracard S, Ducrocq X, et al. Intracranial aneurysms: clinical value of 3D digital subtraction angiography in the therapeutic decision and endovascular treatment. Radiology 2001;218:799–808.[Abstract/Free Full Text]
- McLaughlin RA, Hipwell J, Hawkes DJ, et al. A comparison of intensity-based registration and feature-based registration for neurointerventions. Proc MICCAI 2002;2489:517–24.
- Hipwell JH, Penney GP, Cox TC, et al. 2D–3D intensity based registration of DSA and MRA—a comparison of similarity measures. Proc MICCAI 2002;2489:501–8.
- Masutani Y, Dohi T, Yamane F, Iseki H, Takakura K. Interactive virtualized display system for intravascular neurosurgery. In: CVRMed/MRCAS'97; Grenoble, France. Berlin: Springer-Verlag, 1997: 427–35.
- Ganapathy S. Decomposition of transformation matrices for robot vision. Patt Recog Letts Dec 1984;401–12.
- Penney GP, Weese J, Little JA, Desmedt P, Hill DLG, Hawkes DJ. A comparison of similarity measures for use in 2D–3D medical image registration. IEEE Trans Med Imag 1998;17:586–95.[Medline]
Top
Abstract
Introduction
