- Email: [email protected]
Impact of fiducial arrangement and registration sequence on target accuracy using a phantom frameless stereotactic navigation model
Journal of Clinical Neuroscience 21 (2014) 1976–1980
Contents lists available at ScienceDirect
Journal of Clinical Neuroscience journal homepage: www.elsevier.com/locate/jocn
Technical Note
Impact of fiducial arrangement and registration sequence on target accuracy using a phantom frameless stereotactic navigation model Timothy R. Smith a,⇑, Divakar S. Mithal a, James A. Stadler a, Camelia Asgarian a, Kenji Muro b, Joshua M. Rosenow a a b
Department of Neurological Surgery, Northwestern University Feinberg School of Medicine, 676 N. St. Clair Street, Suite 2210, Chicago, IL 60611, USA Advocate Illinois Masonic Medical Center, Chicago, IL, USA
a r t i c l e
i n f o
Article history: Received 10 February 2014 Accepted 6 April 2014
Keywords: Accuracy Brain mapping Frameless stereotaxy Functional neurosurgery Image-guided neurosurgery Neuronavigation Stereotactic surgery
a b s t r a c t Modern frameless stereotactic techniques utilize scalp fiducial markers for registration. Anecdotal reports from surgeons indicate a variety of methods for improving accuracy using different fiducial arrangements and registration sequences. The few published studies on registration accuracy do not provide a simple and systematic method for determining target accuracy. Nine different arrangements of ten fiducial markers were attached to a model. Ten separate markers were designated as targets for evaluation of registration accuracy. We systematically registered each of the arrangements over multiple trials, in one of four sequences, and then measured the targets. The target coordinates were compared against the established target values, and a root-mean-square deviation (RMSD) was derived. A systematic multivariate analysis determined the effects of different variables on the RMSD. We found no correlation between the ‘‘Registration Accuracy’’ provided by Medtronic (Medtronic Navigation, Louisville, CO, USA) and our RMSD representing targeting accuracy (R = 0.008). RMSD did vary for different fiducial arrangements. We found no significant difference between the various sequences of fiducial arrangement. Thus, regardless of fiducial arrangement, registration sequence has no impact on accuracy. Fiducial arrangements distributed optimally across the skull, however, allowed for significantly improved accuracy. Further studies are required to determine which different arrangements of fiducials are relevant for specific procedures. Published by Elsevier Ltd.
1. Introduction Stereotaxy is a technique by which intracranial areas of interest can be targeted with a high degree of accuracy and precision. The technique was originally conducted with a frame of known dimensions, which served as the reference with which targeting was accomplished. Framed stereotaxy has the advantages of providing a support for placement of targeting devices; however patients often report discomfort, as the frame is usually placed with only local anesthesia [1]. For many intracranial procedures, framed stereotaxy was the standard method for targeting intracranial regions with accuracy, including such procedures as surgery for movement disorders and brain biopsies [2]. Over the last two decades, framed stereotaxy has been largely supplanted by frameless approaches using computerized neuronavigation systems. Frameless stereotaxy minimizes patient
⇑ Corresponding author. Tel.: +1 312 695 0087; fax: +1 312 695 0225. E-mail address: [email protected] (T.R. Smith). http://dx.doi.org/10.1016/j.jocn.2014.04.006 0967-5868/Published by Elsevier Ltd.
discomfort, reduces time in the operating room, and is compatible with various forms of imaging [3]. Furthermore, efficacy may be comparable between frame-based and frameless methods [4–6]. Frameless stereotaxy requires that an image of the patient be mapped to the actual patient in the surgical suite in order to perform image-guided surgery. This is achieved through the registration of fiducials on the patient’s scalp with the location of the same fiducial on an imaging study to allow the neuronavigation system to correlate the locations of each. This ‘‘point-pair’’ method is known to be more accurate than other, less formalized systems [7]. Many approaches have tried to identify the types of error that can occur during frameless registration [8–11]. Specifically, there are three types of error described by Maurer and colleagues [8]: 1. Fiducial localization error indicates a difference between where a fiducial is in the image space and where it is in the patient space. This error can result from movement of fiducials on the skin of a patient between the time of imaging and surgery.
T.R. Smith et al. / Journal of Clinical Neuroscience 21 (2014) 1976–1980
2. Fiducial registration error indicates a measured difference after registration of the fiducial. This error is a reflection of the variability that occurs during registration of the system. 3. Target registration error (TRE) indicates the error in measuring a target once a system is registered. This is the most clinically relevant as it reflects the surgeon’s confidence to rely on the image guided system. Each of these three errors confers different problems, but TRE is most often the subject of debate as it is most likely to be user-dependent. Within the ‘‘point-pair’’ system, there have been anecdotal reports (personal communications) of varying TRE accuracy with different registration techniques, although none has been reported in the literature to our knowledge. West et al. have given guidelines for generating high TRE accuracy by placing greater numbers of fiducials in certain patterns [9]. Wang and Song have provided fiducial templates aimed at reducing TRE based on similar principles [12]. Neither group, however, has analyzed the sequence in which the fiducials are registered, thus leaving one component of the registration method unexamined. Other methods for accurate image-based stereotaxy are based on mathematical models as opposed to empirically generated data [11,13]. Given the importance of TRE minimization, determining methods to reduce TRE are an essential component of improving frameless stereotaxy. We present a study that analyzes different components of TRE accuracy through rigorous testing in a phantom model.
2. Methods Ninety adhesive fiducial markers were firmly attached to a dry skull (Biomet microfixation, Warsaw, IN, USA) model in nine different arrangements (Fig. 1). The nine arrangements of fiducials placed on the skull were termed axial (AX, a band around the head in a coronal plane), bifrontal (BIF, only across the frontal regions), coronal–sagittal (CS, one line in a single sagittal plane, another in a
1977
single coronal plane), parasagittal (PAR, bilateral lines, in parallel sagittal planes), two-axis (TA, two different bands around the head in parallel axial planes), wide-whole (WW, spread across the whole skull), and three different random arrangements (RA1, RA2 and RA3). Ten fiducials were not part of a registration arrangement and served solely as designated targets for evaluation of registration accuracy. The 10 targets were used as a random set to eliminate variability of target distance to fiducial arrangements. It is commonly understood that proximity of fiducial sets to intracranial targets impact registration accuracy, and by randomly assigning targets to each arrangement, and carrying out multiple trials of each arrangement and sequence, we strategically controlled for this variability. A CT scan and MRI of the skull was obtained and the data were uploaded onto a Medtronic Stealth Station Treon Plus Neuronavigation System running Cranial Mach 4 software (Medtronic Navigation, Louisville, CO, USA). The images were merged and the merges were inspected and found to be adequate. Ten target fiducials distributed across the skull were registered and used as control coordinates for all further registration incidents. Each subsequent registration included a measurement of the target fiducials. We systematically registered the nine different fiducial arrangements into image space using a right-to-left (RL) sequence in five trials. In a similar fashion, we then registered three trials each with variation of the fiducial registration sequence: left-right (LR), alternating back-front (BF), and antero-posterior (AP) within each arrangement. Registration and targeting error data was captured by the Stealth Station system, using a proprietary program created by Medtronic Navigation. Root-mean-square deviations (RMSD) were calculated for target fiducials within each combination of fiducial arrangement and registration sequence. Thus, the measured target values were superimposed on the established values, and the errors for each target were condensed into a single RMSD value. For each registration p trial we used the equation: RMSD = (1/N [R {i – N} d {i}2]), where N is the number of targets, i represents the target number (1–10) and d the distance between the established target coordinates
Fig. 1. Photographs with orthogonal views of phantom skull with nine numerically labeled (1–9) fiducial arrangements and one target fiducial arrangement labeled ‘‘XX’’.
1978
T.R. Smith et al. / Journal of Clinical Neuroscience 21 (2014) 1976–1980
and the measured coordinates. The d was measured using x, y and z coordinates and a standard sum of squares method. The resulting RMSD represented a composite measure of error for all 10 target fiducials. We analyzed the data with both analysis of variance (ANOVA), analysis of covariance (ANCOVA) and multiple linear regression (MLR). With an RMSD for each registration trial, we were able to use ANOVA to compare accuracy for different fiducial arrangements and different registration sequences. We then used Tukey post hoc analysis of ordered means if the null-hypothesis was rejected, to determine the source of variation. Tukey’s test allowed ordered comparisons between means, beginning with those having the greatest difference. We used two-way ANOVA to compare the interaction of specific arrangements with specific sequences. We also calculated the relationship of the calculated RMSD and the ‘‘Registration Accuracy’’ number provided by the Medtronic software. We created regression models to determine the impact of fiducial arrangement, registration sequence, and the interaction between arrangement and sequence on accuracy (RMSD), controlling for ‘‘Registration Accuracy’’. All RMSD were calculated using Excel (Microsoft, Seattle, WA, USA). All trial data were entered for systematic statistical analysis using the Statistical Package for the Social Sciences version 21 and Sigma Stat (both SPSS/IBM, Chicago, IL, USA).
random arrangements, RA1, RA2, and RA3). Accounting for all trials of each registration sequence (RL: five trials, AP: three trials, BF: three trials, LR: three trials) resulted in 14 trials to compare for each arrangement. Ultimately, the three random arrangements were analyzed individually, as well as condensed into a single random group. Using ANOVA, we demonstrated that the variation between groups was significantly different to the variation within groups (F = 16.6, p = .001), thus rejecting the null hypothesis that fiducial arrangement has no effect on RMSD (Table 1). In order to determine which arrangements account for rejection of the null hypothesis, we compared Tukey post hoc analyses of the arrangement dataset (Table 1). This allowed for analysis of ordered means, as our primary interest was in determining the arrangement with the lowest RMSD. First we demonstrated that the AX arrangement was significantly worse than TA, WW, BIF, RA2 and CS. CS and RA2 were not significantly different, nor were BIF and WW. By this analysis TA was the only arrangement to consistently give a significantly different RMSD. When the analysis was inverted and measures of central tendency were examined, TA and WW had the lowest average RMSD (TA = 1.35 mm, standard deviation [SD] = .065 mm; WW = 1.43 mm, SD = .061 mm), and both of these were significantly lower than the other arrangements (p = .001).
3. Results
Having noted the impact of fiducial arrangement, we next wanted to determine if the sequence in which fiducials were registered had an independent impact on RMSD. The four registration sequences (RL, LR, BF, and AP) produced no significant difference in RMSD (Table 2, p = .871).
3.1. RMSD correlation with Medtronic StealthStation ‘‘Registration Accuracy’’ We compared our measured RMSD with the published ‘‘Registration Accuracy’’ from the Medtronic StealthStation software. The ‘‘Registration Accuracy’’ of each registration trial was recorded and correlated against our measured RMSD in Figure 2. No correlation was found between our measured RMSD and the published ‘‘Registration Accuracy’’ from Medtronic (R2 = .008). 3.2. Impact of fiducial arrangement on RMSD Nine different arrangements of fiducials were included for registration purposes (AX, BIF, CS, PAR, TA, WW, and three different
3.3. Impact of fiducial registration sequence on RMSD
3.4. Analysis of interactions between arrangement and sequence Despite the lack of statistically significant differences between fiducial registration sequences, we wanted to clarify whether there was an interaction between the fiducial arrangement and registration sequence. This was accomplished with two statistical methods: ANCOVA and MLR. Two-way ANOVA indicated the possibility of an interaction between arrangement and sequence (Table 3, p = .01). Each of the subsequent subgroup analyses looking at variation of sequence within each arrangement, however, demonstrated no effect of registration sequence. Thus the interaction was thought to be due entirely to the arrangement effect seen previously in the bivariate analysis. The ANCOVA was then verified and expanded using multiple linear regression. Using RMSD as the dependent variable, a model was fit with fiducial arrangement as the predictor, controlling for ‘‘Registration Accuracy’’, registration sequence, and the interaction of arrangement and sequence (Table 3). A model was created with all nine arrangement categories, as well as a model with arrangement re-categorized into TA or not. While the table includes the model with arrangement categorized dichotomously, both models produced statistically significant differences. MLR confirmed that controlling for registration sequence, ‘‘Registration Accuracy’’ and the interaction of arrangement and sequence, fiducial arrangement significantly changed the RMSD (p = .001). Of note, the interaction of arrangement and sequence was not found to be significant in the MLR model (p = .162). 4. Discussion
Fig. 2. Correlation of published Medtronic Stealth Registration Accuracy measurement (unknown units) (Medtronic Navigation, Louisville, CO, USA) with root-meansquare error calculation (in millimeters). Linear regression is included with R2 = 0.008.
In this study we sought to analyze the impact of fiducial arrangement and registration sequence in the accuracy of frameless stereotactic targeting. The question of what method
1979
T.R. Smith et al. / Journal of Clinical Neuroscience 21 (2014) 1976–1980 Table 1 One-way analysis of variance for root-mean-square error by fiducial arrangement with Tukey Honestly Significant Difference post hoc analysis for two-axis arrangement RMSD
Sum of squares
df
Mean square
F
p value
1.101 1.315 2.416
6 119 125
.184 .011
16.613
.000*
Post hoca for two-axis
RMSD (mm)
SD (mm)
Axial Bifrontal Coronal–sagittal Parasagittal Two-axis Wide whole Random
1.68 1.48 1.52 1.62 1.35 1.43 1.56
.095 .100 .083 .181 .065 .061 .104
Between groups Within groups Total
.000* .031* .001* .000* .000* .298 .000*
df = degrees of freedom, RMSD = root mean square deviation, SD = standard deviation. * p 6 .05 considered significant. a Tukey Honestly Significant Difference.
Table 2 One-way analysis of variance for root-mean-square error by fiducial registration sequence RMSD Between groupsa Within groupsa Total
Sum of squares
df
Mean square
F
p value
.152 26.02 16.173
3 121 124
.051 .215
.236
.871
a Groups = right-to-left, left-to-right, alternating back-to-front, antero-posterior. df = degrees of freedom, RMSD = root-mean-square deviation.
attains the highest accuracy is important for the best outcomes in neurosurgical procedures. Furthermore, as demonstrated here, the exact meaning of the ‘‘Registration Accuracy’’ term produced by Medtronic software is not readily apparent. Thus our systematic analysis provides some insight for how to approach registration of a frameless stereotaxic system. We chose to assess accuracy using RMSD for two primary reasons. First, our goal was to assess errors by measurement of distance, which is often measured by root-mean-square methods. Second, we had a set of 10 points in three dimensions, which is similar to a rigid-body problem. RMSD is often utilized in matching protein structures, thus making it a valuable tool for comparing multiple points in three-dimensional space [14]. Since the targets were on the skull and their coordinates were known from the imaging studies just like the registration fiducials, the calculation of an RMSD was used as a measure of accuracy. The targeting RMSD is not to be confused with the ‘‘Registration Accuracy’’ published by Medtronic, which is also an RMSD. The
‘‘Registration Accuracy’’ number is derived from comparing the registration fiducials to their expected location from the imaging studies. Thus, the ‘‘Registration Accuracy’’ reading does not given an idea of the accuracy with respect to anatomical targets, and this must be calculated separately [15,16]. Previous work has indicated that the number and distribution of fiducials is a primary factor in reducing TRE [9]. Our data indicate this is likely to be the case. In particular, the most accurate arrangement was found to be TA, characterized by two circumferential bands of fiducials arranged in the parallel axial planes. This arrangement reflected a wide distribution of fiducials with very little redundancy in the x, y, or z planes. This was also reflected, although to a lesser degree, in the WW arrangement, which had fiducials placed evenly over the entirety of the skull. By contrast, all arrangements with a linear base (AX, PAR, CS, BIF) were less accurate, likely due to their redundancy in one or more of the x, y, and z planes. Not surprisingly, the three RA arrangements, which had random placement of fiducials, were highly variable in their accuracy, likely reflecting the different characteristics of these arrangements. We conclude that the lower RMSD we measured were related to an optimized distribution of fiducials for our target distribution. There have been anecdotal accounts that different registration sequences could have an impact on registration accuracy. To rule out this possibility, we tested a number of different sequences including LR, RL, BF, and AP. These different sequences demonstrated no independent or interactive impact on RMSD, indicating than registration sequences are not likely to play a role in minimizing TRE.
Table 3 Two-way analysis of variance for root-mean-square error by fiducial arrangement and multiple linear regression model for root mean square error by arrangement controlling for StealthStation ‘‘Registration Accuracy’’, fiducial arrangement, registration sequence and interaction RMSD Arrangement Sequence Arrangement Sequence Within groups Total MLR model factorsa Fiducial arrangement Registration sequence StealthStation ‘‘Registration Accuracy’’b Interactionc *
Sum of squares
df
Mean square
F
p value
1.20452 .03817 .40706 .76588 2.41562
8 3 24 90 125
.15056 .01272 .01696 .00851 .01932
17.69 1.49 1.99
<.001* 0.221 .011*
Beta .432 .020 .046 .155
95% CI .266 .028 .256 .010
p 6 .05 considered significant. Dependent variable = root mean square deviation. b Medtronic Navigation, Louisville, CO, USA. C Interaction of fiducial arrangement and registration sequence; forced entry (one step), F-statistic = 8.668, df = 4, p < .001. CI = confidence interval, df = degrees of freedom, MLR = multiple linear regression, RMSD = root mean square deviation. a
p value .104 .023 .149 .002
.000* .850 .605 .162
1980
T.R. Smith et al. / Journal of Clinical Neuroscience 21 (2014) 1976–1980
One major limitation of the work presented here is that we used multiple targets on the surface of the skull. Most surgeries will require a more specific target located somewhere within the skull. This target will not be able to have a fiducial placed directly on it, as we did for our targets. We believe that further studies are warranted using a phantom model similar to the one we have presented here. Wang and Song have proposed methods and templates for fiducial placement to reduce TRE [12,17,18]. Their methods of interactive fiducial placement are likely too difficult to implement intra-operatively, and their fiducial templates, while an improvement over the arbitrary placements that are currently used, are still limited. Our data support their conclusion that optimal fiducial placement will result in decreased TRE, and that optimal fiducial placement is dependent on the lesion location. We believe, however, that using the approach we have detailed here is a simple, efficient and accurate method to develop fiducial arrangements optimized for any intracranial lesion. 5. Conclusion Using a dry skull model and a Medtronic StealthStation, we analyzed the efficacy of different arrangements and registration sequences for targeting accuracy. We have demonstrated that of nine unique fiducial arrangements, one in particular was consistently the most accurate. Furthermore, we have demonstrated that the sequence in which the arrangement is registered is of no consequence. These findings indicate that surgeons need not have a specific sequence in their fiducial registration technique. Rather, an important aspect of frameless stereotaxy appears to be the arrangement of fiducials for specific surgical indications. To this end, further studies are indicated regarding which arrangements are ideal for individual applications of image-guided frameless stereotaxy. Conflicts of Interest/Disclosures The authors declare that they have no financial or other conflicts of interest in relation to this research and its publication.
References [1] Eggers G, Mühling J, Marmulla R. Image-to-patient registration techniques in head surgery. Int J Oral Maxillofac Surg 2006;35:1081–95. [2] Kitchen ND, Lemieux L, Thomas DG. Accuracy in frame-based and frameless stereotaxy. Stereotact Funct Neurosurg 1993;61:195–206. [3] Dorward NL, Alberti O, Dijkstra A, et al. Clinical introduction of an adjustable rigid instrument holder for frameless stereotactic interventions. Comput Aided Surg 1997;2:180–5. [4] Quiñones-Hinojosa A, Ware ML, Sanai N, et al. Assessment of image guided accuracy in a skull model: comparison of frameless stereotaxy techniques vs. frame-based localization. J Neurooncol 2006;76:65–70. [5] Woodworth GF, McGirt MJ, Samdani A, et al. Frameless image-guided stereotactic brain biopsy procedure: diagnostic yield, surgical morbidity, and comparison with the frame-based technique. J Neurosurg 2006;104:233–7. [6] Murphy MA, McKenzie RL, Kormos DW, et al. Frameless stereotaxis for the insertion of lumbar pedicle screws. J Clin Neurosci 1994;1:257–60. [7] Woerdeman PA, Willems PW, Noordmans HJ, et al. Application accuracy in frameless image-guided neurosurgery: a comparison study of three patient-toimage registration methods. J Neurosurg 2007;106:1012–6. [8] Maurer CR Jr, Fitzpatrick JM, Wang MY, et al. Registration of head volume images using implantable fiducial markers. IEEE Trans Med Imaging 1997;16: 447–62. [9] West JB, Fitzpatrick JM, Toms SA, et al. Fiducial point placement and the accuracy of point-based, rigid body registration. Neurosurgery 2001;48:810–6 [discussion 816–7]. [10] Liu H, Yu Y, Schell MC, et al. Optimal marker placement in photogrammetry patient positioning system. Med Phys 2003;30:103–10. [11] Fitzpatrick JM, West JB. The distribution of target registration error in rigidbody point-based registration. IEEE Trans Med Imaging 2001;20:917–27. [12] Wang M, Song Z. Distribution templates of the fiducial points in image-guided neurosurgery. Neurosurgery 2010;66:143–50 [discussion 150–1]. [13] Atuegwu NC, Galloway RL. Volumetric characterization of the Aurora magnetic tracker system for image-guided transorbital endoscopic procedures. Phys Med Biol 2008;53:4355–68. [14] Jewett AI, Huang CC, Ferrin TE. MINRMS: an efficient algorithm for determining protein structure similarity using root-mean-squared-distance. Bioinformatics 2003;19:625–34. [15] Gralla J, Guzman R, Brekenfeld C, et al. High-resolution three-dimensional T2weighted sequence for neuronavigation: a new setup and clinical trial. J Neurosurg 2005;102:658–63. [16] Sweeney RA, Bale RJ, Moncayo R, et al. Multimodality cranial image fusion using external markers applied via a vacuum mouthpiece and a case report. Strahlenther Onkol 2003;179(4):254–60. [17] Wang M, Song Z. Improving target registration accuracy in image-guided neurosurgery by optimizing the distribution of fiducial points. Int J Med Robot 2009;5:26–31. [18] Wang M, Song Z. Guidelines for the placement of fiducial points in imageguided neurosurgery. Int J Med Robot 2010;6:142–9.
Contents lists available at ScienceDirect
Journal of Clinical Neuroscience journal homepage: www.elsevier.com/locate/jocn
Technical Note
Impact of fiducial arrangement and registration sequence on target accuracy using a phantom frameless stereotactic navigation model Timothy R. Smith a,⇑, Divakar S. Mithal a, James A. Stadler a, Camelia Asgarian a, Kenji Muro b, Joshua M. Rosenow a a b
Department of Neurological Surgery, Northwestern University Feinberg School of Medicine, 676 N. St. Clair Street, Suite 2210, Chicago, IL 60611, USA Advocate Illinois Masonic Medical Center, Chicago, IL, USA
a r t i c l e
i n f o
Article history: Received 10 February 2014 Accepted 6 April 2014
Keywords: Accuracy Brain mapping Frameless stereotaxy Functional neurosurgery Image-guided neurosurgery Neuronavigation Stereotactic surgery
a b s t r a c t Modern frameless stereotactic techniques utilize scalp fiducial markers for registration. Anecdotal reports from surgeons indicate a variety of methods for improving accuracy using different fiducial arrangements and registration sequences. The few published studies on registration accuracy do not provide a simple and systematic method for determining target accuracy. Nine different arrangements of ten fiducial markers were attached to a model. Ten separate markers were designated as targets for evaluation of registration accuracy. We systematically registered each of the arrangements over multiple trials, in one of four sequences, and then measured the targets. The target coordinates were compared against the established target values, and a root-mean-square deviation (RMSD) was derived. A systematic multivariate analysis determined the effects of different variables on the RMSD. We found no correlation between the ‘‘Registration Accuracy’’ provided by Medtronic (Medtronic Navigation, Louisville, CO, USA) and our RMSD representing targeting accuracy (R = 0.008). RMSD did vary for different fiducial arrangements. We found no significant difference between the various sequences of fiducial arrangement. Thus, regardless of fiducial arrangement, registration sequence has no impact on accuracy. Fiducial arrangements distributed optimally across the skull, however, allowed for significantly improved accuracy. Further studies are required to determine which different arrangements of fiducials are relevant for specific procedures. Published by Elsevier Ltd.
1. Introduction Stereotaxy is a technique by which intracranial areas of interest can be targeted with a high degree of accuracy and precision. The technique was originally conducted with a frame of known dimensions, which served as the reference with which targeting was accomplished. Framed stereotaxy has the advantages of providing a support for placement of targeting devices; however patients often report discomfort, as the frame is usually placed with only local anesthesia [1]. For many intracranial procedures, framed stereotaxy was the standard method for targeting intracranial regions with accuracy, including such procedures as surgery for movement disorders and brain biopsies [2]. Over the last two decades, framed stereotaxy has been largely supplanted by frameless approaches using computerized neuronavigation systems. Frameless stereotaxy minimizes patient
⇑ Corresponding author. Tel.: +1 312 695 0087; fax: +1 312 695 0225. E-mail address: [email protected] (T.R. Smith). http://dx.doi.org/10.1016/j.jocn.2014.04.006 0967-5868/Published by Elsevier Ltd.
discomfort, reduces time in the operating room, and is compatible with various forms of imaging [3]. Furthermore, efficacy may be comparable between frame-based and frameless methods [4–6]. Frameless stereotaxy requires that an image of the patient be mapped to the actual patient in the surgical suite in order to perform image-guided surgery. This is achieved through the registration of fiducials on the patient’s scalp with the location of the same fiducial on an imaging study to allow the neuronavigation system to correlate the locations of each. This ‘‘point-pair’’ method is known to be more accurate than other, less formalized systems [7]. Many approaches have tried to identify the types of error that can occur during frameless registration [8–11]. Specifically, there are three types of error described by Maurer and colleagues [8]: 1. Fiducial localization error indicates a difference between where a fiducial is in the image space and where it is in the patient space. This error can result from movement of fiducials on the skin of a patient between the time of imaging and surgery.
T.R. Smith et al. / Journal of Clinical Neuroscience 21 (2014) 1976–1980
2. Fiducial registration error indicates a measured difference after registration of the fiducial. This error is a reflection of the variability that occurs during registration of the system. 3. Target registration error (TRE) indicates the error in measuring a target once a system is registered. This is the most clinically relevant as it reflects the surgeon’s confidence to rely on the image guided system. Each of these three errors confers different problems, but TRE is most often the subject of debate as it is most likely to be user-dependent. Within the ‘‘point-pair’’ system, there have been anecdotal reports (personal communications) of varying TRE accuracy with different registration techniques, although none has been reported in the literature to our knowledge. West et al. have given guidelines for generating high TRE accuracy by placing greater numbers of fiducials in certain patterns [9]. Wang and Song have provided fiducial templates aimed at reducing TRE based on similar principles [12]. Neither group, however, has analyzed the sequence in which the fiducials are registered, thus leaving one component of the registration method unexamined. Other methods for accurate image-based stereotaxy are based on mathematical models as opposed to empirically generated data [11,13]. Given the importance of TRE minimization, determining methods to reduce TRE are an essential component of improving frameless stereotaxy. We present a study that analyzes different components of TRE accuracy through rigorous testing in a phantom model.
2. Methods Ninety adhesive fiducial markers were firmly attached to a dry skull (Biomet microfixation, Warsaw, IN, USA) model in nine different arrangements (Fig. 1). The nine arrangements of fiducials placed on the skull were termed axial (AX, a band around the head in a coronal plane), bifrontal (BIF, only across the frontal regions), coronal–sagittal (CS, one line in a single sagittal plane, another in a
1977
single coronal plane), parasagittal (PAR, bilateral lines, in parallel sagittal planes), two-axis (TA, two different bands around the head in parallel axial planes), wide-whole (WW, spread across the whole skull), and three different random arrangements (RA1, RA2 and RA3). Ten fiducials were not part of a registration arrangement and served solely as designated targets for evaluation of registration accuracy. The 10 targets were used as a random set to eliminate variability of target distance to fiducial arrangements. It is commonly understood that proximity of fiducial sets to intracranial targets impact registration accuracy, and by randomly assigning targets to each arrangement, and carrying out multiple trials of each arrangement and sequence, we strategically controlled for this variability. A CT scan and MRI of the skull was obtained and the data were uploaded onto a Medtronic Stealth Station Treon Plus Neuronavigation System running Cranial Mach 4 software (Medtronic Navigation, Louisville, CO, USA). The images were merged and the merges were inspected and found to be adequate. Ten target fiducials distributed across the skull were registered and used as control coordinates for all further registration incidents. Each subsequent registration included a measurement of the target fiducials. We systematically registered the nine different fiducial arrangements into image space using a right-to-left (RL) sequence in five trials. In a similar fashion, we then registered three trials each with variation of the fiducial registration sequence: left-right (LR), alternating back-front (BF), and antero-posterior (AP) within each arrangement. Registration and targeting error data was captured by the Stealth Station system, using a proprietary program created by Medtronic Navigation. Root-mean-square deviations (RMSD) were calculated for target fiducials within each combination of fiducial arrangement and registration sequence. Thus, the measured target values were superimposed on the established values, and the errors for each target were condensed into a single RMSD value. For each registration p trial we used the equation: RMSD = (1/N [R {i – N} d {i}2]), where N is the number of targets, i represents the target number (1–10) and d the distance between the established target coordinates
Fig. 1. Photographs with orthogonal views of phantom skull with nine numerically labeled (1–9) fiducial arrangements and one target fiducial arrangement labeled ‘‘XX’’.
1978
T.R. Smith et al. / Journal of Clinical Neuroscience 21 (2014) 1976–1980
and the measured coordinates. The d was measured using x, y and z coordinates and a standard sum of squares method. The resulting RMSD represented a composite measure of error for all 10 target fiducials. We analyzed the data with both analysis of variance (ANOVA), analysis of covariance (ANCOVA) and multiple linear regression (MLR). With an RMSD for each registration trial, we were able to use ANOVA to compare accuracy for different fiducial arrangements and different registration sequences. We then used Tukey post hoc analysis of ordered means if the null-hypothesis was rejected, to determine the source of variation. Tukey’s test allowed ordered comparisons between means, beginning with those having the greatest difference. We used two-way ANOVA to compare the interaction of specific arrangements with specific sequences. We also calculated the relationship of the calculated RMSD and the ‘‘Registration Accuracy’’ number provided by the Medtronic software. We created regression models to determine the impact of fiducial arrangement, registration sequence, and the interaction between arrangement and sequence on accuracy (RMSD), controlling for ‘‘Registration Accuracy’’. All RMSD were calculated using Excel (Microsoft, Seattle, WA, USA). All trial data were entered for systematic statistical analysis using the Statistical Package for the Social Sciences version 21 and Sigma Stat (both SPSS/IBM, Chicago, IL, USA).
random arrangements, RA1, RA2, and RA3). Accounting for all trials of each registration sequence (RL: five trials, AP: three trials, BF: three trials, LR: three trials) resulted in 14 trials to compare for each arrangement. Ultimately, the three random arrangements were analyzed individually, as well as condensed into a single random group. Using ANOVA, we demonstrated that the variation between groups was significantly different to the variation within groups (F = 16.6, p = .001), thus rejecting the null hypothesis that fiducial arrangement has no effect on RMSD (Table 1). In order to determine which arrangements account for rejection of the null hypothesis, we compared Tukey post hoc analyses of the arrangement dataset (Table 1). This allowed for analysis of ordered means, as our primary interest was in determining the arrangement with the lowest RMSD. First we demonstrated that the AX arrangement was significantly worse than TA, WW, BIF, RA2 and CS. CS and RA2 were not significantly different, nor were BIF and WW. By this analysis TA was the only arrangement to consistently give a significantly different RMSD. When the analysis was inverted and measures of central tendency were examined, TA and WW had the lowest average RMSD (TA = 1.35 mm, standard deviation [SD] = .065 mm; WW = 1.43 mm, SD = .061 mm), and both of these were significantly lower than the other arrangements (p = .001).
3. Results
Having noted the impact of fiducial arrangement, we next wanted to determine if the sequence in which fiducials were registered had an independent impact on RMSD. The four registration sequences (RL, LR, BF, and AP) produced no significant difference in RMSD (Table 2, p = .871).
3.1. RMSD correlation with Medtronic StealthStation ‘‘Registration Accuracy’’ We compared our measured RMSD with the published ‘‘Registration Accuracy’’ from the Medtronic StealthStation software. The ‘‘Registration Accuracy’’ of each registration trial was recorded and correlated against our measured RMSD in Figure 2. No correlation was found between our measured RMSD and the published ‘‘Registration Accuracy’’ from Medtronic (R2 = .008). 3.2. Impact of fiducial arrangement on RMSD Nine different arrangements of fiducials were included for registration purposes (AX, BIF, CS, PAR, TA, WW, and three different
3.3. Impact of fiducial registration sequence on RMSD
3.4. Analysis of interactions between arrangement and sequence Despite the lack of statistically significant differences between fiducial registration sequences, we wanted to clarify whether there was an interaction between the fiducial arrangement and registration sequence. This was accomplished with two statistical methods: ANCOVA and MLR. Two-way ANOVA indicated the possibility of an interaction between arrangement and sequence (Table 3, p = .01). Each of the subsequent subgroup analyses looking at variation of sequence within each arrangement, however, demonstrated no effect of registration sequence. Thus the interaction was thought to be due entirely to the arrangement effect seen previously in the bivariate analysis. The ANCOVA was then verified and expanded using multiple linear regression. Using RMSD as the dependent variable, a model was fit with fiducial arrangement as the predictor, controlling for ‘‘Registration Accuracy’’, registration sequence, and the interaction of arrangement and sequence (Table 3). A model was created with all nine arrangement categories, as well as a model with arrangement re-categorized into TA or not. While the table includes the model with arrangement categorized dichotomously, both models produced statistically significant differences. MLR confirmed that controlling for registration sequence, ‘‘Registration Accuracy’’ and the interaction of arrangement and sequence, fiducial arrangement significantly changed the RMSD (p = .001). Of note, the interaction of arrangement and sequence was not found to be significant in the MLR model (p = .162). 4. Discussion
Fig. 2. Correlation of published Medtronic Stealth Registration Accuracy measurement (unknown units) (Medtronic Navigation, Louisville, CO, USA) with root-meansquare error calculation (in millimeters). Linear regression is included with R2 = 0.008.
In this study we sought to analyze the impact of fiducial arrangement and registration sequence in the accuracy of frameless stereotactic targeting. The question of what method
1979
T.R. Smith et al. / Journal of Clinical Neuroscience 21 (2014) 1976–1980 Table 1 One-way analysis of variance for root-mean-square error by fiducial arrangement with Tukey Honestly Significant Difference post hoc analysis for two-axis arrangement RMSD
Sum of squares
df
Mean square
F
p value
1.101 1.315 2.416
6 119 125
.184 .011
16.613
.000*
Post hoca for two-axis
RMSD (mm)
SD (mm)
Axial Bifrontal Coronal–sagittal Parasagittal Two-axis Wide whole Random
1.68 1.48 1.52 1.62 1.35 1.43 1.56
.095 .100 .083 .181 .065 .061 .104
Between groups Within groups Total
.000* .031* .001* .000* .000* .298 .000*
df = degrees of freedom, RMSD = root mean square deviation, SD = standard deviation. * p 6 .05 considered significant. a Tukey Honestly Significant Difference.
Table 2 One-way analysis of variance for root-mean-square error by fiducial registration sequence RMSD Between groupsa Within groupsa Total
Sum of squares
df
Mean square
F
p value
.152 26.02 16.173
3 121 124
.051 .215
.236
.871
a Groups = right-to-left, left-to-right, alternating back-to-front, antero-posterior. df = degrees of freedom, RMSD = root-mean-square deviation.
attains the highest accuracy is important for the best outcomes in neurosurgical procedures. Furthermore, as demonstrated here, the exact meaning of the ‘‘Registration Accuracy’’ term produced by Medtronic software is not readily apparent. Thus our systematic analysis provides some insight for how to approach registration of a frameless stereotaxic system. We chose to assess accuracy using RMSD for two primary reasons. First, our goal was to assess errors by measurement of distance, which is often measured by root-mean-square methods. Second, we had a set of 10 points in three dimensions, which is similar to a rigid-body problem. RMSD is often utilized in matching protein structures, thus making it a valuable tool for comparing multiple points in three-dimensional space [14]. Since the targets were on the skull and their coordinates were known from the imaging studies just like the registration fiducials, the calculation of an RMSD was used as a measure of accuracy. The targeting RMSD is not to be confused with the ‘‘Registration Accuracy’’ published by Medtronic, which is also an RMSD. The
‘‘Registration Accuracy’’ number is derived from comparing the registration fiducials to their expected location from the imaging studies. Thus, the ‘‘Registration Accuracy’’ reading does not given an idea of the accuracy with respect to anatomical targets, and this must be calculated separately [15,16]. Previous work has indicated that the number and distribution of fiducials is a primary factor in reducing TRE [9]. Our data indicate this is likely to be the case. In particular, the most accurate arrangement was found to be TA, characterized by two circumferential bands of fiducials arranged in the parallel axial planes. This arrangement reflected a wide distribution of fiducials with very little redundancy in the x, y, or z planes. This was also reflected, although to a lesser degree, in the WW arrangement, which had fiducials placed evenly over the entirety of the skull. By contrast, all arrangements with a linear base (AX, PAR, CS, BIF) were less accurate, likely due to their redundancy in one or more of the x, y, and z planes. Not surprisingly, the three RA arrangements, which had random placement of fiducials, were highly variable in their accuracy, likely reflecting the different characteristics of these arrangements. We conclude that the lower RMSD we measured were related to an optimized distribution of fiducials for our target distribution. There have been anecdotal accounts that different registration sequences could have an impact on registration accuracy. To rule out this possibility, we tested a number of different sequences including LR, RL, BF, and AP. These different sequences demonstrated no independent or interactive impact on RMSD, indicating than registration sequences are not likely to play a role in minimizing TRE.
Table 3 Two-way analysis of variance for root-mean-square error by fiducial arrangement and multiple linear regression model for root mean square error by arrangement controlling for StealthStation ‘‘Registration Accuracy’’, fiducial arrangement, registration sequence and interaction RMSD Arrangement Sequence Arrangement Sequence Within groups Total MLR model factorsa Fiducial arrangement Registration sequence StealthStation ‘‘Registration Accuracy’’b Interactionc *
Sum of squares
df
Mean square
F
p value
1.20452 .03817 .40706 .76588 2.41562
8 3 24 90 125
.15056 .01272 .01696 .00851 .01932
17.69 1.49 1.99
<.001* 0.221 .011*
Beta .432 .020 .046 .155
95% CI .266 .028 .256 .010
p 6 .05 considered significant. Dependent variable = root mean square deviation. b Medtronic Navigation, Louisville, CO, USA. C Interaction of fiducial arrangement and registration sequence; forced entry (one step), F-statistic = 8.668, df = 4, p < .001. CI = confidence interval, df = degrees of freedom, MLR = multiple linear regression, RMSD = root mean square deviation. a
p value .104 .023 .149 .002
.000* .850 .605 .162
1980
T.R. Smith et al. / Journal of Clinical Neuroscience 21 (2014) 1976–1980
One major limitation of the work presented here is that we used multiple targets on the surface of the skull. Most surgeries will require a more specific target located somewhere within the skull. This target will not be able to have a fiducial placed directly on it, as we did for our targets. We believe that further studies are warranted using a phantom model similar to the one we have presented here. Wang and Song have proposed methods and templates for fiducial placement to reduce TRE [12,17,18]. Their methods of interactive fiducial placement are likely too difficult to implement intra-operatively, and their fiducial templates, while an improvement over the arbitrary placements that are currently used, are still limited. Our data support their conclusion that optimal fiducial placement will result in decreased TRE, and that optimal fiducial placement is dependent on the lesion location. We believe, however, that using the approach we have detailed here is a simple, efficient and accurate method to develop fiducial arrangements optimized for any intracranial lesion. 5. Conclusion Using a dry skull model and a Medtronic StealthStation, we analyzed the efficacy of different arrangements and registration sequences for targeting accuracy. We have demonstrated that of nine unique fiducial arrangements, one in particular was consistently the most accurate. Furthermore, we have demonstrated that the sequence in which the arrangement is registered is of no consequence. These findings indicate that surgeons need not have a specific sequence in their fiducial registration technique. Rather, an important aspect of frameless stereotaxy appears to be the arrangement of fiducials for specific surgical indications. To this end, further studies are indicated regarding which arrangements are ideal for individual applications of image-guided frameless stereotaxy. Conflicts of Interest/Disclosures The authors declare that they have no financial or other conflicts of interest in relation to this research and its publication.
References [1] Eggers G, Mühling J, Marmulla R. Image-to-patient registration techniques in head surgery. Int J Oral Maxillofac Surg 2006;35:1081–95. [2] Kitchen ND, Lemieux L, Thomas DG. Accuracy in frame-based and frameless stereotaxy. Stereotact Funct Neurosurg 1993;61:195–206. [3] Dorward NL, Alberti O, Dijkstra A, et al. Clinical introduction of an adjustable rigid instrument holder for frameless stereotactic interventions. Comput Aided Surg 1997;2:180–5. [4] Quiñones-Hinojosa A, Ware ML, Sanai N, et al. Assessment of image guided accuracy in a skull model: comparison of frameless stereotaxy techniques vs. frame-based localization. J Neurooncol 2006;76:65–70. [5] Woodworth GF, McGirt MJ, Samdani A, et al. Frameless image-guided stereotactic brain biopsy procedure: diagnostic yield, surgical morbidity, and comparison with the frame-based technique. J Neurosurg 2006;104:233–7. [6] Murphy MA, McKenzie RL, Kormos DW, et al. Frameless stereotaxis for the insertion of lumbar pedicle screws. J Clin Neurosci 1994;1:257–60. [7] Woerdeman PA, Willems PW, Noordmans HJ, et al. Application accuracy in frameless image-guided neurosurgery: a comparison study of three patient-toimage registration methods. J Neurosurg 2007;106:1012–6. [8] Maurer CR Jr, Fitzpatrick JM, Wang MY, et al. Registration of head volume images using implantable fiducial markers. IEEE Trans Med Imaging 1997;16: 447–62. [9] West JB, Fitzpatrick JM, Toms SA, et al. Fiducial point placement and the accuracy of point-based, rigid body registration. Neurosurgery 2001;48:810–6 [discussion 816–7]. [10] Liu H, Yu Y, Schell MC, et al. Optimal marker placement in photogrammetry patient positioning system. Med Phys 2003;30:103–10. [11] Fitzpatrick JM, West JB. The distribution of target registration error in rigidbody point-based registration. IEEE Trans Med Imaging 2001;20:917–27. [12] Wang M, Song Z. Distribution templates of the fiducial points in image-guided neurosurgery. Neurosurgery 2010;66:143–50 [discussion 150–1]. [13] Atuegwu NC, Galloway RL. Volumetric characterization of the Aurora magnetic tracker system for image-guided transorbital endoscopic procedures. Phys Med Biol 2008;53:4355–68. [14] Jewett AI, Huang CC, Ferrin TE. MINRMS: an efficient algorithm for determining protein structure similarity using root-mean-squared-distance. Bioinformatics 2003;19:625–34. [15] Gralla J, Guzman R, Brekenfeld C, et al. High-resolution three-dimensional T2weighted sequence for neuronavigation: a new setup and clinical trial. J Neurosurg 2005;102:658–63. [16] Sweeney RA, Bale RJ, Moncayo R, et al. Multimodality cranial image fusion using external markers applied via a vacuum mouthpiece and a case report. Strahlenther Onkol 2003;179(4):254–60. [17] Wang M, Song Z. Improving target registration accuracy in image-guided neurosurgery by optimizing the distribution of fiducial points. Int J Med Robot 2009;5:26–31. [18] Wang M, Song Z. Guidelines for the placement of fiducial points in imageguided neurosurgery. Int J Med Robot 2010;6:142–9.