Download j48-2009-jns-neurosurgery

J Neurosurg 111:1201–1206, 2009
Surface-based facial scan registration in neuronavigation
procedures: a clinical study
Clinical article
Reuben R. Shamir, M.Sc.,1 Moti Freiman, M.Sc.,1 Leo Joskowicz, Ph.D.,1
Sergey Spektor, M.D., Ph.D., 2 and Yigal Shoshan, M.D. 2
School of Engineering and Computer Science; and 2Department of Neurosurgery, Hadassah Medical Center,
Hebrew University, Jerusalem, Israel
1
Object. Surface-based registration (SBR) with facial surface scans has been proposed as an alternative for the
commonly used fiducial-based registration in image-guided neurosurgery. Recent studies comparing the accuracy of
SBR and fiducial-based registration have been based on a few targets located on the head surface rather than inside
the brain and have yielded contradictory conclusions. Moreover, no visual feedback is provided with either method
to inform the surgeon about the estimated target registration error (TRE) at various target locations. The goals in the
present study were: 1) to quantify the SBR error in a clinical setup, 2) to estimate the targeting error for many target
locations inside the brain, and 3) to create a map of the estimated TRE values superimposed on a patient’s head image.
Methods. The authors randomly selected 12 patients (8 supine and 4 in a lateral position) who underwent neurosurgery with a commercial navigation system. Intraoperatively, scans of the patients’ faces were acquired using a fast
3D surface scanner and aligned with their preoperative MR or CT head image. In the laboratory, the SBR accuracy
was measured on the facial zone and estimated at various intracranial target locations. Contours related to different
TREs were superimposed on the patient’s head image and informed the surgeon about the expected anisotropic error
distribution.
Results. The mean surface registration error in the face zone was 0.9 ± 0.35 mm. The mean estimated TREs for
targets located 60, 105, and 150 mm from the facial surface were 2.0, 3.2, and 4.5 mm, respectively. There was no
difference in the estimated TRE between the lateral and supine positions. The entire registration procedure, including
positioning of the scanner, surface data acquisition, and the registration computation usually required < 5 minutes.
Conclusions. Surface-based registration accuracy is better in the face and frontal zones, and error increases as
the target location lies further from the face. Visualization of the anisotropic TRE distribution may help the surgeon
to make clinical decisions. The observed and estimated accuracies and the intraoperative registration time show that
SBR using the fast surface scanner is practical and feasible in a clinical setup. (DOI: 10.3171/2009.3.JNS081457)
Key Words • image-guided therapy • surface-based registration • registration error
I
mage-guided surgery based on preoperative images
and intraoperative navigation has become the standard of care for many neurosurgical procedures. A
key step in image-guided surgery is the accurate intraoperative alignment, commonly called “registration,”
be­tween preoperative MR images or CT scans and the
in­tra­operative patient situation.
Three registration approaches are currently available
to achieve proper alignment: 1) FBR, 2) IBR, and 3) SBR.
Fiducial-based registration relies on anatomical landmarks
Abbreviations used in this paper: FBR = fiducial-based registration; IBR = image-based registration; SBR = surface-based registration; SRE = surface registration error; TRE = target registration
error.
J Neurosurg / Volume 111 / December 2009
and/or fiducial markers affixed to the patient’s skin or skull
prior to scanning. Intraoperatively, the surgeon touches
the fiducial markers and/or anatomical landmarks with
a tracked probe and pairs these points with their counterparts in the preoperative image.3,12,18,21 Image-based
registration uses intraoperatively acquired images, such
as ultrasonographic or MR images, instead of landmarks
to establish image-to-patient alignment.1,20 Surface-based
reg­istration uses intraoperative 3D surface points from the
patient’s anatomy, such as the face and forehead, to compute the image-to-patient transformation.4,6,8
This article contains some figures that are displayed in color
on­line but in black and white in the print edition.
1201
R. R. Shamir et al.
Surface-based registration has the following advantages over FBR and IBR: 1) it is marker free; 2) it is fast,
intuitive, and easy to use; 3) it is not subject to manual
pairing and localization errors; and 4) its accuracy is surgeon independent given that the surface is automatically
generated by the scanner.
Authors of recent studies have compared the accuracy of FBR and landmark-based registration in a clinical
environment.4,7,9–11,13–16,21,22 Their conclusions are inconsistent: some authors reported that SBR is significantly less
accurate than FBR,7,15,16,21,22 whereas others showed that
SBR’s accuracy is similar to that of FBR.4,9–11,13,14
A possible explanation for the different results is the
variation in the selected registration fiducial and target
locations. As shown in clinical studies, the fiducial setup
and target location do influence targeting accuracy.17–19
Therefore, any direct comparison of the FBR and SBR
methods may be biased to the specific fiducial and target
locations selected in the study. Furthermore, researchers
in all of these studies have measured the targeting error
for only a few targets located on the head surface, and thus
they have made incomplete error assessments, especially
for targets inside the brain. Since the actual target locations inside the brain are unknown, an error estimation
method must be used to evaluate the localization error.
In the study on FBR by West et al.,19 patient head
images were augmented by overlaid contours that showed
the estimated TREs. This visualization provided additional clinically relevant information that could help a
neurosurgeon in the decision-making process.
The goals in the present study on SBR were: 1) to
quantify the SBR error in a clinical setup, 2) to estimate
the targeting error for various locations inside the brain,
and 3) to help neurosurgeons better assess the TRE by
visualizing it on a patient’s head image.
Methods
We randomly selected 12 patients who were scheduled to undergo image-guided neurosurgeries for various
clinical indications. Patient consent was obtained in all
cases.
Data Acquisition Protocol
All patients underwent imaging the day before surgery.
Contrast-enhanced T1-weighted MR images with a 1-mm
slice width (Signa, GE Medical Systems) were obtained in
11 patients (93%). Contrast CT head scans (helical Twin
Flash scanner, Philips Medical Systems) with a 1.3-mm
slice width were obtained in 1 patient (7%). Magnetic resonance images showed 256 × 256 × 200 voxels, with a voxel
size of 0.93 × 0.93 × 1 mm; CT scans showed 512 × 512 ×
100 voxels, with a voxel size similar to that used in the MR
imaging. The imaging volume included areas of interest
inside the brain and on the facial surface.
Before surgery, the head surface was automatically
segmented and reconstructed, as reported in our previous
study.8 Seventy-three thousand to 115,000 facial surface
points were extracted from the MR imaging data sets, and
64,000 from the CT data set. After imaging and preoperative planning, the patients were taken to the operating
1202
Fig. 1. Photograph of the operating room setup showing the patient
and surface scanner positions.
room. After general anesthesia had been induced, 8 patients were placed supine and 4 in a lateral position. Because nasogastric tube fixation to the nose and eye draping can significantly alter facial surface geometry around
the eyes and nose, we performed surface scanning prior
to this time.
One to 5 scans of each of the 12 patients’ faces were
acquired with a root mean square point location accuracy
of 0.3 ± 0.2 mm by using a commercial 3D optical surface
scanner (faceSCAN II, Breuckmann) in the setup shown
in Fig. 1. We acquired 34 different intraoperative scans
in the 12 patients. For each patient, several scans were
obtained using different scanner positions and environmental setups. We recorded all the data and brought it to
our laboratory for further study.
Experimental Protocol
We selected the anterior facial region from the preoperative surface scan and semiautomatically extracted
points on the patient’s upper facial surface, including the
forehead, eyes, and nose, by using custom software. The
preoperative MR image or CT scan was then aligned to
an intraoperative surface scan using a robust 2-step registration algorithm (Fig. 2). The first step established a
coarse correspondence based on 4 eye and nose bridge
landmarks that are automatically extracted from both the
MR image/CT scan and the facial surface scan data. The
second step derived the registration transformation by
aligning the point clouds from the MR image/CT scan
and the facial surface with the robust iterative closest
point registration method.2
J Neurosurg / Volume 111 / December 2009
Surface-based registration in neurosurgery procedures
operating room illumination, distance and orientation of the
surface scanner with respect to the patient, and the tools,
tubes, and other equipment around the patient’s head.
Results
Fig. 2. Image demonstrating the face surface points from an intraoperative surface scan (gray points) registered to forefront surface points
extracted from a preoperative MR image (red/dark points).
To quantify the clinical accuracy of the SBR, we computed 2 error measurements: 1) the actual SRE, and 2) the
estimated TRE. The actual SRE is the root-mean-square
distance between facial surface points on the intraoperative facial surface scan and their closest corresponding
points on the MR image/CT head surface after registration. Note that when the registration is based on fiducials,
a compatible error is known as the fiducial registration
error. The actual TRE is the distance between the true location of a predefined target and its predicted location as
computed with SBR. Since targets are located inside the
brain, their true location and actual TRE cannot be measured. Instead, we used the analytical formula described
by Fitzpatrick et al.5 to estimate the TRE of various target locations inside the brain. This method was originally
designed for FBR under the assumption of independent,
identically and normally distributed landmark localization errors. From a careful examination of this method
we concluded that it could be used for SBR under similar
assumptions.
Ten targets were selected on each patient image at
increasing distances from the facial surface, and the corresponding TRE estimations were computed. Based on
these estimations, isoerror value contours were superimposed on a representative MR image, providing an estimated TRE map of the brain. The results were tabulated
separately for the supine and lateral positions.
We also investigated registration sensitivity to scanning conditions, including internal surface scanner parameters, patient anatomical properties, and environmental
parameters. Scanner parameters included the intensity of
projected structured light and expected object brightness.
Patient parameters included skin color and the presence/
absence of facial hair. Environmental parameters included
J Neurosurg / Volume 111 / December 2009
The intraoperative surface scans were registered to the
MR imaging/CT data sets in < 1 minute on a standard personal computer (2.4-GHz processor, Pentium 4, Intel).
Results, which were tabulated separately for the supine
or lateral positions, are summarized in Table 1. There was
no difference in the measured SREs and TREs in the 2
groups. The mean SRE over all 12 patients was 0.91 ± 0.35
mm, with a range of 0.69–1.31 mm. The mean estimated
TRE increased as the target location lay further from the
face surface scan (Fig. 3). For example, targets located at
60, 105, and 150 mm from the patient’s face had estimated
TRE ranges of 1.53–2.84, 2.26–4.53, and 2.99–6.40 mm,
respectively.
To better understand and visualize the clinical relevance of our results, we created estimated TRE maps by
selecting a representative MR imaging slice and drawing
on it the estimated TREs as isoerror value lines superimposed on the image (Fig. 4).
An important finding in our study is the poor correlation between the SRE and the estimated TRE. In particular, we observed that close SRE values do not necessarily
predict close TRE values: Data Sets 9 and 11 have nearly
identical SRE values (0.98 ± 0.38 mm and 0.99 ± 0.42
mm, respectively) but increasingly different TRE values
(2.69 and 2.17 mm at 60 mm, 4.52 and 3.49 mm at 105
mm, and 6.40 and 4.88 mm at 150 mm, respectively). Our
recent research corroborates these findings.17,18
We also noted that to obtain the best results from the
surface scan data, the surface scanner should be placed at
a distance of 0.9–1.2 m above the patient’s head, with the
internal scanner parameters for light intensity set to the
75–100 U range and the expected scanned object parameter set to “dark.” Factors such as patient skin color, facial
hair, and adjacent surgical instruments did not affect registration accuracy.
Discussion
Our results indicated that SBR has a low average SRE
of 0.91 mm in the facial zone. These results are excellent
considering the intrinsic image resolution error and are
comparable to those obtained using FBR based on comparable MR imaging/CT data sets.3,12,18,19 The average estimated TRE increases as the target location lies further
away from the face, with errors of 4 mm in deep targets
located 105 mm from the face.
Observed differences between the SRE and estimated
TRE in SBR were also noteworthy. The SRE, which is
the only accuracy measure provided by commercial neuronavigation software for both registration methods (SBR
and FBR), underestimates the TRE for structures inside the
brain and thus can be misleading. As our data show, the
TRE of SBR for targets at a distance of > 60 mm can be
relatively high and sometimes unacceptable, even in cases
in which the SRE is low. The discrepancy may be crucial
for small deep targets in, for example, the basal ganglia,
1203
R. R. Shamir et al.
TABLE 1: Summary of SBR accuracy results in 12 patients*
Case No.†
1
2
3
4
5
6
7§
8
average for supine position
9
10
11
12
average for lat position
overall average
Average No. of Face Surface
No. of
Points (in thousands)
Intraop
Surface
Extracted
Scan
From Surface
Extracted
FRE
Data Sets
Scans
From MRI/CT Measured (mm)
5
4
5
2
3
3
3
2
—
2
2
2
1
—
—
1.7
2.8
1.8
2.1
2.0
1.5
2.4
2.1
2.1
1.7
2.1
2.2
2.9
2.1
2.1
81
90
110
100
90
94
64
115
93
73
96
100
104
92
93
0.69 ± 0.23
1.08 ± 0.42
0.87 ± 0.33
0.76 ± 0.30
0.81 ± 0.30
0.83 ± 0.28
1.31 ± 0.55
0.85 ± 0.31
0.91 ± 0.25
0.98 ± 0.38
0.90 ± 0.34
0.99 ± 0.42
1.09 ± 0.44
0.95 ± 0.15
0.91 ± 0.35
TRE Estimated (mm)
60 mm‡
105 mm‡
150 mm‡
1.58
2.59
3.63
2.18
3.49
4.86
1.85
3.10
4.39
1.53
2.40
3.32
1.88
3.04
4.25
2.10
3.46
4.85
2.84
4.53
6.29
1.80
2.85
3.95
1.99
3.23
4.52
2.69
4.52
6.40
1.87
2.99
4.16
2.17
3.49
4.88
1.60
2.26
2.99
2.05
3.30
4.60
2.00 ± 0.55 3.24 ± 0.95 4.53 ± 1.41
* Values represent the means ± SDs, unless indicated otherwise. Patients in the first 8 cases were supine, whereas those in the
last 4 cases were placed in the lateral position. — = not applicable.
† Also referred to as data set number.
‡ Distance between the target and the patient’s face.
§ Patient underwent preoperative CT scanning; all other patients underwent preoperative MR imaging.
pineal region, and cerebral peduncle as well as the upper
brainstem when it is approached from anterior to the coronal suture. Therefore, a surgeon using SBR should consider
the SRE as an incomplete error assessment.
Our estimated TRE map (Fig. 4) shows the expected
TRE at various zones of the brain image. This type of input can help the neurosurgeon select his or her approach
(entry point and trajectory) using neuronavigation to a
specific target while estimating the success rate (for example, correct diagnosis on stereotactic biopsy) and complication probability.
Our data also suggested that SBR is feasible in a
busy operating room routine. The surface scanner is used
before the skin incision, does not require direct contact
or sterilization, and can be moved out of the operating
field for the rest of the surgery. It requires < 5 minutes to
set up, with the actual scanning taking only a few seconds. The surface scan-to-image data set registration
process is fully automatic and relatively short, requiring
< 2 minutes’ computation time. In comparison, standard
neuronavigation systems with FBR require the surgeon
to manually locate and touch the anatomical landmarks,
which can be an inaccurate and time-consuming process,
especially when performed by an inexperienced surgeon
or when the patient is placed in the lateral position.
Another commercially available alternative is a tracked
laser-beam line scanner (Z-touch, BrainLab). To acquire
surface points, the surgeon points the laser beam to the patient’s face and moves the beam in a predefined pattern to
1204
record several hundred face surface points. This process is
time consuming, user dependent, and error prone. In contrast, the fast 3D surface scanner automatically acquires
tens of thousands of points, which can increase registration
Fig. 3. Graph revealing the mean (center line) and SD (interval) of
the estimated TRE with respect to the location of the target. It is shows
that as the target lies closer to the face, the SBR is more accurate and
the TRE is decreasing.
J Neurosurg / Volume 111 / December 2009
Surface-based registration in neurosurgery procedures
a miniature robot for keyhole neurosurgery.8 We envisage
using this method for other image-guided surgery applications for which real-time tracking is not available.
Conclusions
Fig. 4. Expected TRE map showing the mean estimated TRE as
isovalue lines superimposed on a representative MR imaging slice.
accuracy, and the surface is acquired in a few seconds. The
registration result is surgeon independent as he or she has
no effect on the generated surface.
The main drawback of our proposed method is the
actual cost of a 3D surface scanner (~ €20,000, although
the prices are dropping) with an accuracy of ± 0.3 mm
or better. In addition, a custom stand to hold the scanner
may be required in the operating room.
Another limitation of our method is that none of the
existing TRE estimation methods are clinically validated
for image-to-patient SBR. Current TRE estimations and
appropriate visualization methods need further investigation and development.
A key limitation of SBR is that it does not include
location information from other parts of the head and
thus can be of limited use for targets in the deep temporal, parietal, and occipital lobes, basal ganglia, pineal
region, and upper brainstem. We envisage 2 possible solutions. The first is the addition of 1 or 2 skin fiducial
markers in strategic head locations. Although the surgeon will need to touch these fiducials with a tracked
probe intraoperatively, the localization uncertainty will
be much smaller than with anatomical landmarks.18 Furthermore, our and other studies indicated that adding a
single fiducial marker in an optimal location can reduce
the TRE by one-half.17,19 A second possible solution is to
use additional uni- or bilateral ear scans. The outer ear
surface provides lateral localization information that can
be added to the frontal surface scan. Although the ear is
flexible and deformable, we have observed that its outer
surface and location with respect to the head is relatively
stable. Thus, the outer ear surface can be extracted from
the preoperative CT/MR image and matched with an intraoperative side surface scan. The conjunction of both
data sets may provide a better distribution of registration
points and allow us to obtain a smaller TRE. For the supine position, bilateral external ear scanning may provide
a combination of 2 opposite surface registration data sets
that may enhance TREs. Further studies will be required
to explore this idea.
We have integrated SBR into a system consisting of
J Neurosurg / Volume 111 / December 2009
A surface-based facial scanning–to–preoperative MR
imaging/CT data registration process is automatic, short,
and feasible with good operating room integration. Our
results on 12 patients showed that the surface registration
error is ~ 1 mm in the facial zone for both supine and lateral patient positions. This level of accuracy is comparable
to that obtained using FBR with skin fiducial markers in
commercially available optical-based neuronavigation systems. The SBR accuracy was better in the face and frontal zones, and the estimated registration error increased as
the target location lay further away from the upper facial
surface and deeper inside the brain. Visualization of the
mean estimated TRE as isovalue lines superimposed on
a patient’s MR image can assist the surgeon in decision
making while planning and executing his neuronavigation
procedures.
Disclosure
This research was supported in part by Magneton Grant No.
34377 from the Israeli Ministry of Industry and Trade.
Acknowledgment
The authors thank Mrs. Shifra Fraifeld at Hadassah-Hebrew
University Medical Center for her editorial assistance in the preparation of this manuscript.
References
1. Amstutz C, Caversaccio M, Kowal J, Bachler R, Nolte LP,
Haus­ler R, et al: A-mode ultrasound-based registration in
com­puter-aided surgery of the skull. Arch Otolaryngol Head
Neck Surg 129:1310–1316, 2003
2. Besl PJ, McKay ND: A method for registration of 3D shapes.
IEEE Trans Pattern Anal Mach Intell 14:239–256, 1992
3. Bjartmarz H, Rehncrona S: Comparison of accuracy and precision between frame-based and frameless stereotactic navigation for deep brain stimulation electrode implantation. Stereotact Funct Neurosurg 85:235–242, 2007
4. Cao A, Thompson RC, Dumpuri P, Dawant BM, Galloway RL,
Ding S, et al: Laser range scanning for image-guided neurosurgery: investigation of image-to-physical space registrations.
Med Phys 35:1593–1605, 2008
5. Fitzpatrick JM, West JB, Maurer CR Jr: Predicting error in
rigid-body point-based registration. IEEE Trans Med Imaging 17:694–702, 1998
6. Grimson WL, Ettinger GJ, White SJ, Lozano-Perez T, Wells
WM, Kikinis R: An automatic registration method for frameless stereotaxy, image guided surgery, and enhanced reality
visualization. IEEE Trans Med Imaging 15:129–140, 1996
7. Hoffmann J, Westendorff C, Leitner C, Bartz D, Reinert S:
Validation of 3D-laser surface registration for image-guided
cranio-maxillofacial surgery. J Craniomaxillofac Surg 33:
13–18, 2005
8. Joskowicz L, Shamir R, Freiman M, Shoham M, Zehavi E,
Umansky F, et al: Image-guided system with miniature robot
for precise positioning and targeting in keyhole neurosurgery.
Comput Aided Surg 11:181–193, 2006
9. Marmulla R, Hassfeld S, Luth T, Mende U, Muhling J: Soft tis-
1205
R. R. Shamir et al.
sue scanning for patient registration in image-guided surgery.
Comput Aided Surg 8:70–81, 2003
10. Marmulla R, Muhling J, Wirtz CR, Hassfeld S: High-resolution
laser surface scanning for patient registration in cranial computer-assisted surgery. Minim Invasive Neurosurg 47:72–78,
2004
11. Mascott CR, Sol JC, Bousquet P, Lagarrigue J, Lazorthes Y,
Lauwers-Cances V: Quantification of true in vivo (application) accuracy in cranial image-guided surgery: influence of
mode of patient registration. Neurosurgery 59 (1 Suppl 1):
ONS146–ONS156, 2006
12. Maurer CR Jr, Fitzpatrick JM, Wang MY, Galloway RL Jr,
Maciunas RJ, Allen GS: Registration of head volume images
using implantable fiducial markers. IEEE Trans Med Imaging 16:447–462, 1997
13. Miga MI, Sinha TK, Cash DM, Galloway RL, Weil RJ: Cortical surface registration for image-guided neurosurgery using
laser-range scanning. IEEE Trans Med Imaging 22:973–
985, 2003
14. Raabe A, Krishnan R, Wolff R, Hermann E, Zimmermann M,
Seifert V: Laser surface scanning for patient registration in intracranial image-guided surgery. Neurosurgery 50:797–803,
2002
15. Schicho K, Figl M, Seemann R, Donat M, Pretterklieber ML,
Birkfellner W, et al: Comparison of laser surface scanning
and fiducial marker-based registration in frameless stereotaxy. Technical note. J Neurosurg 106:704–709, 2007
16. Schlaier J, Warnat J, Brawanski A: Registration accuracy and
practicability of laser-directed surface matching. Comput Aided Surg 7:284–290, 2002
17. Shamir RR, Joskowicz L, Spektor S, Shoshan Y: Localization
and registration accuracy in image guided neurosurgery: a
clinical study. Int J CARS 4:45–52, 2009
1206
18. Shamir RR, Joskowicz L, Shoshan Y: Optimal landmarks
selection and fiducial marker placement for minimal target
registration error in image-guided neurosurgery. Proc SPIE
7261:[epub ahead of print], 2009
19. West JB, Fitzpatrick JM, Toms SA, Maurer CR Jr, Maciunas
RJ: Fiducial point placement and the accuracy of point-based,
rigid body registration. Neurosurgery 48:810–817, 2001
20. Wirtz CR, Tronnier VM, Bonsanto MM, Knauth M, Staubert
A, Albert FK, et al: Image-guided neurosurgery with intraoperative MRI: update of frameless stereotaxy and radicality
control. Stereotact Funct Neurosurg 68:39–43, 1997
21. Woerdeman PA, Willems PWA, Noordmans HJ, Berkelbach
van der Sprenkel JW: Clinical accuracy of neuronavigation
using registration methods based on point-pairs or surface
matching. Int J CARS 1:297–299, 2006 (Abstract)
22. Woerdeman PA, Willems PW, Noordmans HJ, Tulleken CA,
Berkelbach van der Sprenkel JW: Application accuracy in
frameless image-guided neurosurgery: a comparison study
of three patient-to-image registration methods. J Neurosurg
106:1012–1016, 2007
Manuscript submitted November 9, 2008.
Accepted March 19, 2009.
Please include this information when citing this paper: published
online April 24, 2009; DOI: 10.3171/2009.3.JNS081457.
Portions of this work have been presented as an extended abstract
at the CARS 2007, Computer-Assisted Radiology and Surgery 21st
International Congress in Berlin, Germany.
Address correspondence to: Reuben R. Shamir, M.Sc., School of
Engineering and Computer Science, Hebrew University, Givat Ram
Campus, Jerusalem, Israel 91904. email: [email protected].
J Neurosurg / Volume 111 / December 2009