A Comparison of 2D-3D Intensity-Based

Registration and Feature-Based Registration

for Neurointerventions

Robert A. McLaughlin1 , John Hipwell2 ,

David J. Hawkes2 , J. Alison Noble1 , James V. Byrne3 , and Tim Cox4
1
Medical Vision Laboratory, Dept. Engineering Science, University of Oxford,
Oxford, England
[email protected]
2
CISG, Division of Radiological Sciences, Guy’s Hospital
King’s College London, London, England
[email protected]
3
Department of Radiology, Radcliffe Infirmary, Oxford, England
4
National Hospital for Neurology and Neurosurgery
Queen Square, London, England

Abstract. Registration of 2D-3D data can improve visualisation during

minimally-invasive neurointerventions. Using four clinical data sets, we
quantitatively compared two approaches: an intensity-based algorithm
and a feature-based algorithm. The intensity-based approach was found
to be more accurate, with an average registration accuracy of 1.4mm,
compared to the feature-based algorithm with an average accuracy of
2.3mm. The intensity-based algorithm was also found to be more reliable.
Reliability of the feature-based algorithm was found to be more sensitive
to the complexity of the vasculature structure.

1 Introduction

The registration of 2D-3D data sets is important in minimally invasive neuro-

interventions, such as the coiling of brain aneurysms or glueing of arteriovenous
malformations (AVM). During such interventions a neuro-radiologist guides a
catheter through the brain vasculature using 2D X-ray images. The 2D nature
of the images can make it difficult to navigate and position the catheter accu-
rately in a complicated 3D angioarchitecture.
One solution would be to utilise a pre-operative phase contrast magnetic
resonance angiography (PC-MRA) scan. Such a scan could be segmented [1][2]
to produce a 3D model of the vasculature. By registering the intra-operative
X-ray image with this 3D model, it would be possible to accurately display the
position of the catheter relative to the 3D model.
In this paper, we compare two approaches to 2D-3D registration: an intensity-
based method [3] and a feature-based method [4]. We compare accuracy and
robustness of these two algorithms on four clinical data sets.

T. Dohi and R. Kikinis (Eds.): MICCAI 2002, LNCS 2489, pp. 517–524, 2002.


c Springer-Verlag Berlin Heidelberg 2002
518 R.A. McLaughlin et al.

2 Method
2.1 Intensity-Based Registration
The intensity-based registration algorithm builds on the work in [3] and itera-
tively optimises the six rigid-body parameters describing the location and rota-
tion of the 3D model. Digitally reconstructed radiographs (DRR) are generated
by casting rays through the segmented volume, and are compared to the digital
subtraction angiography (DSA) image using the gradient difference similarity
measure [5]. Gradient images are computed for both the the DSA image and
the DRR using 3x3 Sobel templates. The gradient difference similarity measure
minimises the difference between these gradient images. Details are given in [3].
Some modifications to the algorithm of [3] were required to adapt it to work
with segmented 3D data and DSA images, rather than unsegmented CT data and
fluoroscopy images. The primary modification was the use of a spherical volume-
of-interest (VOI), manually defined around the feature of interest (aneurysm or
AVM). Only voxels lying within the VOI were used in the registration.
The VOI was projected onto the DRR as a circular mask. A concentric cir-
cular mask with one quarter the radius was then defined, and pixels within this
smaller mask were used in an initial registration. The radius of the smaller mask
was then doubled and this larger mask was used to refine the registration. The
centre of rotation for the volume was set to be the centre of the VOI.
To reduce processing time at each stage, a multi-resolution strategy was
adopted whereby the DRRs and DSA images were sub-sampled by a factor of
four. These dimensions were subsequently doubled until the optimisation of the
parameters was completed with both images at their full resolutions.

2.2 Feature-Based Registration

The feature-based registration algorithm [4] first skeletonises vessels in the DSA
image, reducing the thickness of each to a single pixel. Blood vessels in the 3D
model are also skeletonised by extracting the medial axis of each vessel. The
algorithm registers the data sets by matching the skeletonised DSA image with
a projection of the skeletonised 3D model. For each 3D point, the closest corre-
sponding point in the skeletonised DSA image is found using a territory-based
correspondence search as described in [4]. Using these pairs of points and the
method outlined in [6], the algorithm finds the optimal rotation and translation
of the 3D model to achieve a registration. The registration was performed in three
stages, using the VOI defined in Section 2.1. Registration was initially performed
with the small mask, refined using the larger mask and finally completed using
the entire DSA image. A similar use of masks was presented in [7].

3 Experiments
3.1 Data
Phase-contrast MRA (PC-MRA) scans were obtained for three patients with
aneurysms (patients 1-3) and one patient with an AVM (patient 4). Scans were
 2D-3D Intensity-Based Registration versus Feature-Based Registration 519

acquired on a Siemens Magnetom Vision 1.5T with voxel size 0.78 x 0.78 x 1.5mm
and image dimensions 256 x 256 x 64. The scans for patients 1, 2 and 4 contained
flow speed information and were segmented as described in [1]. An improved scan
was used for patient 3, giving both flow speed and flow direction information,
and this extra information was used to give an improved segmentation [2]. Vi-
sualisations of the segmented MRA data sets are shown in Figure 1.

Patient 1 Patient 2

Patient 3 Patient 4

Fig. 1. 3D visualisations of segmented MRA scans.

For each patient, two DSA runs at different orientations were acquired using
a GE Medical Systems Advantx DX, and digitised from the PAL composite
video signal at an image resolution of 512 × 512 pixels using a Matrox Meteor
II framegrabber. For each DSA run, three to seven images were acquired at
half second intervals. These were used to generate two images: a maximal image
where the images were combined so that the maximal level of contrast over
the run is recorded for each pixel; and a single-frame image, where an image
that had maximal opacification of the arterial system was chosen. A distortion-
correction phantom and software were used to correct for pincushion distortion
in the images [5]. Typical DSA images produced are shown in Figure 2.

3.2 Calculation of “Gold-Standard” Registration

Parameters for the gold-standard registration may be described as either intrin-

sic or extrinsic. Intrinsic parameters describe properties of the imaging system,
520 R.A. McLaughlin et al.

Patient 1 Patient 2 Patient 3 Patient 4

Fig. 2. The first of two DSA runs obtained for each patient. Patients 1 and 2 show
single frame DSA images, while Patents 3 and 4 show maximal DSA images.

such as the perspective projection matrix. Extrinsic parameters describe the

orientation (rotation) and position (translation) of the 3D model [5]
Intrinsic parameters for the gold standard registration were computed from
parameters obtained from the X-ray machine display during acquisition. As no
fiducial markers were available in either the PC-MRA scans or the DSA images,
the extrinsic parameters were obtained by a manual registration performed by
JVB (neuro-radiologist). The manual registration was performed using 3D vi-
sualisation software which simulated X-ray images for a specified translation
and rotation, allowing the neuro-radiologist to align clinically relevant points
in the images. To test the stability of these manual results, registrations for pa-
tient 2: DSA run 1 and patient 4: DSA run 1 were each repeated eight times, and
variation in the results were computed using the reprojection distance described
in the next section.

3.3 Experiments for Accuracy and Robustness

The segmented MRA data sets were registered with both maximal and single-
frame DSA images. Starting positions for the registrations were chosen by per-
turbing the gold standard values by set amounts. This methodology was used
in [5]. Four experiments were performed, with the amount of perturbation in-
creased each time, as shown in Table 1. For each experiment, different combi-
nations of the four perturbations resulted in sixteen different starting positions.
Note that there were no in-plane translations (δX or δY ), as these can be ac-
curately calculated by selecting a single corresponding point in both the DSA
image and the DRR simulated from the MRA data.
To measure accuracy, the reprojection distance was used, as defined by Ma-
sutani et al. [8]. A number of anatomically visible points on the segmented 3D
model were chosen, along with the corresponding points in the DSA image. Using
the rotation and translation matrix resulting from each registration, the position
of the 3D points was recomputed. The minimum distance (in mm) from each
point to the ray passing from the X-ray source to the corresponding DSA im-
age point was then calculated. This gave a measurement of the accuracy of the
registration when projecting from 3D to 2D. A discussion of the measurement
 2D-3D Intensity-Based Registration versus Feature-Based Registration 521

Table 1. Perturbations of the starting positions from the gold standard for four of the
six rigid-body parameters.

Experiment # δZ δθx δθy δθz

1 ±25 mm ±4◦ ±4◦ ±4◦
2 ±50 mm ±8◦ ±8◦ ±8◦
3 ±75 mm ±12◦ ±12◦ ±12◦
4 ±100 mm ±16◦ ±16◦ ±16◦

can be found in [5]. Finally, the average RMS error of all such points for each
experiment was computed. If the average RMS error for a particular registration
was less than 4 mm, the registration was judged to have succeeded.

4 Experiment and Results

Figure 3a plots registration accuracy for each algorithm, using both the maximal
DSA images and the single-frame DSA images. These are plotted against the
variability in the manual ’gold-standard’ registration, which was computed as
1.7mm. Figure 3b plots the percentage of successful registrations. Only successful
registrations were used in computing the accuracies shown in figure 3a. Results
of typical registrations are displayed in figure 4.

3
Feature: single frame DSA
2.5 Feature: maximal DSA
RMS error (mm)

Intensity: single frame DSA

2
Intensity: maximal DSA
1.5 Manual registration

0.5

0
1 2 3 4
Experiment number

a.
100
90
Feature: single frame DSA
80 Feature: maximal DSA
70 Intensity: single frame DSA
% successes

60 Intensity: maximal DSA

50
40
30
20
10
0
1 2 3 4
Experiment number

b.

Fig. 3. (a) Registration accuracy. (b) Registration reliability.

The percentage of successful registrations varied with the data sets, being
notably higher with Patients 1 and 2 than with Patients 3 and 4. Graphs showing
the percentage of successful registrations with each data set are shown in figure 5.
522 R.A. McLaughlin et al.

a. b. c. d.

e. f.

Fig. 4. Results of registration. Registered 3D vessels are overlaid in black. (a) Original
DSA image for Patient 2. (b) Typical successful registration for patient 2. (c, d) Failed
registrations for patient 2. (e) Original DSA image for Patient 4. (f) Typical successful
registration for patient 4.

100
Patient 1
90
80 Patient 2
70 Patient 3
% successes

60 Patient 4
50
40
30
20
10
0
1 2 3 4
Experiment number

a.
100
90
Patient 1
80 Patient 2
70 Patient 3
% successes

60 Patient 4
50
40
30
20
10
0
1 2 3 4
Experiment number

b.

Fig. 5. Registration reliability for each patient data set. (a) Feature-based algorithm.
(b) Intensity-based algorithm.

5 Discussion

The intensity-based algorithm had the greater accuracy of the two algorithms,
with an average accuracy of 1.4mm. This compared to an average value of 2.3mm
 2D-3D Intensity-Based Registration versus Feature-Based Registration 523

for the feature-based algorithm. Recall that the feature-based algorithm registers
a skeleton of the 3D model with a skeleton of the DSA image. It is thus sensitive
to inaccuracies in the position of the 2D and 3D skeletonised points. This ac-
counts for its lower accuracy when compared to the intensity-based algorithm,
which registers the intensity values of every individual pixel. Note that the dif-
ference in image quality between the maximal and single-frame DSA images did
not noticeably alter the accuracies for either algorithm.
The intensity-based algorithm was also more robust. This is in contrast to
the experimental results of [7], which found the feature-based algorithm to be
more robust. The experiments in [7] were performed using the far simpler vas-
culature of an in-vitro silicon aneurysm phantom (middle cerebral artery bifur-
cation aneurysm). Our results suggest that while for simple angioarchitectures
the feature-based approach may be more robust, in complicated situations the
intensity-based approach is superior. The results in figure 5 support this conclu-
sion. The robustness trends shown in the graphs fall into two distinct classes,
with Patients 1 and 2 proving to be more robust than Patients 3 and 4. Recall
that while the scans for Patients 1 and 2 contained only flow speed informa-
tion, Patient 3 contained both flow speed and direction information. This led
to a more complicated segmentation, with small vessels detected. Patient 4 was
complex due to the angioarchitecture of the AVM. These results suggest that
robustness of the feature-based approach could be greatly improved if, in some
initial stage of processing, the vasculature in the 3D model and DSA image could
be simplified to contain only the most significant vessels.
An essential difference between the two algorithms lies in the method by
which they combine conflicting information and iteratively improve the current
state of the registration. The intensity-based algorithm tests each minor pertur-
bation to the current rotation and translation, minimising a similarity measure
that is summed over the entire data set. In contrast, in the feature-based al-
gorithm each pair of matching points (one from the 3D skeleton and one from
the skeletonised DSA image) specify an optimal change to the present rotation
and translation. It is the average rotation and translation that is chosen. This
method renders the algorithm sensitive to a misregistration of one or two erro-
neous vessels, as these will produce greatly different estimates for the rotation
and translation. This suggests that the use of a robust fitting method such as
RANSAC [9] may greatly improve the reliability of the feature-based algorithm.
The computation time of the algorithms has important ramifications for the
clinical suitability of either approach to registration. The feature-based algorithm
is far less computationally intensive than the intensity-based algorithm, result-
ing in a much faster registration. This is because the feature-based algorithm
operates on a small number of skeletonised points, rather than the exhaustive
pixel-based approach of the intensity-based algorithm. In future work, we will
seek to quantify these differences.
524 R.A. McLaughlin et al.

6 Conclusion
We have compared an intensity-based and a feature-based registration algorithm
for the registration of 3D PC-MRA data to DSA images. The algorithms were
tested using four clinical PC-MRA data sets and eight DSA runs. The intensity-
based registration algorithm produced more accurate registrations, with an av-
erage RMS reprojection error of 1.4 mm. The feature-based algorithm was found
to have an average RMS reprojection error of 2.3 mm.
The intensity-based algorithm was found to converge to the correct solution
with greater reliability. Our results suggest that reliability of the feature-based
algorithm are more effected by the complexity of the angioarchitecture than is
the intensity-based method. In future work we will explore whether the feature-
based approach may be made more reliable by the incorporation of a robust
fitting method.

Acknowledgements
We wish to thank Dr. G. P. Penney, K. Rhode, Dr. A.C.S. Chung and Dr. Y.
Kita for their help in undertaking this research. This work was supported by
ESPRC grants GR/M55008 and GR/M55015.

References
1. Chung, A.C.S., Noble, J.A.: Statistical 3D vessel segmentation using a Rician dis-
tribution. In: Proc. MICCAI. (1999) 82–89
2. Chung, A., Noble, J.: Fusing magnitude and phase information for vascular seg-
mentation in phase contrast mr angiograms. In: Proc. MICCAI. (2000) 166–175
3. Penney, G.P., Weese, J., Little, J.A., Desmedt, P., Hill, D.L.G., Hawkes, D.J.: A
comparison of similarity measures for use in 2d-3d medical image registration. IEEE
Transactions on Medical Imaging 17 (1998) 586–595
4. Kita, Y., Wilson, D.L., Noble, J.A.: Real-time registration of 3D cerebral vessels to
X-ray angiograms. In: MICCAI’98. (1998) 1125–1133
5. Penney, G.P.: Registration of Tomographic Images to X-ray Projections for Use in
Image Guided Interventions. Phd thesis, University College London, CISG, Division
of Radiological Sciences, Guy’s Hospital, King’s College London, London SE1 9RT
England (2000)
6. Heuring, J.J., Murray, D.W.: Visual head tracking and slaving for visual telepres-
ence. In: Proc. of IEEE Int. Conf. on Robotics and Automation. (1996) 2908–2914
7. McLaughlin, R.A., Hipwell, J., Penney, G.P., Rhode, K., Chung, A., Noble, J.A.,
Hawkes, D.J.: Intensity-based registration versus feature-based registration for
neurointerventions. In: Proceedings of Medical Image Understanding and Analy-
sis (MIUA). (2001) 69–72
8. Masutani, Y., Dohi, T., et al., F.Y.: Interactive virtualized display system for
intravascular neurosurgery. In: CVRMed-MRCAS’97. (1997) 427–435
9. Fischler, M.A., Bolles, R.C.: Random sample consensus: A paradigm for model
fitting with applications to image analysis and automated cartography. Communi-
cations of the ACM 24 (1981) 381–395