Introduction

Optical neuronavigation systems have achieved widespread acceptance, and manufacturers claim reliability and accuracy. Nevertheless, the technical complexity of a navigational environment requires experimental evaluation of different types of errors, knowledge of registration techniques, awareness of the device’s limitations and reliability of accuracy feedback. Surgeons often use the square root of the mean squared deviation of registration (RMS), which is calculated and displayed by most systems, as accuracy feedback. However, some authors do not regard the registration error calculated this way as a trustworthy measure of accuracy [14, 41]. The initiative for this study arose from the observation that despite good values given from the system, such as mean distance and RMS, decreased accuracy may occur after surface registration.

In the literature, there are only few experimental phantom studies with the Stryker (Stryker Navigation, Leibinger, Freiburg, Germany) and the BrainLAB VectorVision (BrainLAB, Feldkirchen-Munich, Germany) systems. In this study, we seek to evaluate and compare the accuracy of these two commercially available systems and quantify clinically relevant influencing factors. Most related studies are evaluations of registration accuracy [8] or evaluate navigation accuracy on the surface alone but not of deep lying target points [41]. Accuracy data on BrainLAB are already reported [1, 8, 9, 20, 31, 40]; however most authors refer to an estimation of clinical accuracy rather than conducting experimental quantitative studies. Accuracy evaluations of the Stryker system are rare [34]. A comparative study of these two systems, which represent two different types of systems, active and passive, has not been published so far. The development of a phantom, simulating skin and soft tissue motility and viscoelastic deformation properties is an innovation, allowing testing experimentally the influence of surface registration and fiducial movement, as well as the influence of soft tissue behaviour when touched by a pointing instrument and during image acquisition. To the best of our knowledge, there is no other literature report of an anthropomorphic head phantom with similar properties of soft tissue simulation for the specific research question [46].

Materials and methods

Stryker® (Stryker Navigation, Leibinger, Freiburg, Germany) is an active optical system consisting of a workstation with planning software, an infrared camera frame with optical sensors, a Patient Tracker and a Pointer with light emitting diodes (LED) (Fig. 1) [34]. VectorVision® Sky (BrainLAB, Feldkirchen-Munich, Germany) is a passive reflective marker system (Fig. 2), consisting of a navigation and a planning workstation, an infrared camera and a touch screen monitor, both attached to the ceiling, a reference star with three reflecting spheres, a Pointer with two reflecting spheres and an alternative laser instrument (z-touch) for surface registration.

Fig. 1
figure 1

Stryker–Leibinger in experimental setup: a workstation with camera and monitor, b pointer, c patient tracker and a standard Plexiglas phantom while targeting with the pointer (phantom is not fixed for phototographic purposes)

Full size image
Fig. 2
figure 2

BrainLAB VectorVision Sky in experimental setup: a standard Plexiglas phantom with the reference star up close on the left, b the BrainLAB monitor and camera attached to the OR ceiling (view through an intraoperative CT gantry)

Full size image

A new phantom was developed for the study of surface registration accuracy and error sources. A transparent cranial model was fitted with eight inner target bushings from the Leibinger Cranial Marker System (Fig. 3), which could get fitted with radiopaque spheres for CT, hollow spheres with gadolinium for MRI, or titanium cylinder caps with a mark for navigation measurements. Two additional cylindrical markers were screwed into the lateral surface as superficial targets. Interchangeable markers with central pivots corresponding exactly to the centres of the imaging markers improve the accurate placement of navigation probes [45–47]. In general, artificial targets allow for a more accurate and precise target definition [2, 28, 46]. Through repeated casting, an outer layer was formed. A silicon gel (Sylgard 527 A & B, Dow Corning, UK) was chosen, which has already been used as a soft tissue simulant for brain [6], breast tissue [3] and—in combination with gelatine—for traumatic deformations [11]. Its simulation advantages are viscoelastic behaviour [6], temperature resistance, allowing for controlled movement, and good magnetic properties similar to the ones of human tissue in T1- and T2-weighted MRI sequences [3]. On top of this layer, a transparent skin-friendly material with good MRI signal from gelatine and glycerine was cast (ProsMaster KI, Kerling International GmbH, Backnang, Germany). The use of a multilayer model affects contact-pressure distribution and agrees with in vivo data regarding deformation of the skin surface [10].

Fig. 3
figure 3

Cranial Marker System set from Leibinger (Freiburg, Germany) used for the targets of the newly developed phantom: a the whole set with screw driver, forceps, drill bit and (from left to right) bushings, titanium cylinders with target point (black), hollow spheres for contrast medium, radiopaque spheres and screw extensions (note the schematic drawing on top of each element), b hollow spheres and black titanium markers in magnification with their dimensions, c bushing (1), hollow sphere for contrast medium (2), titanium marker (3), drill bit (4) and screw extension (5)

Full size image

Imaging protocols

The phantom was scanned with a helical cone beam CT (Volume Zoom, Siemens) with no gantry tilt, slice thickness of 2.0 mm recalculated to 1.0 mm, 512 × 512 matrix and in-plane pixel size of 0.43 × 0.43 mm2, as well as with an MR tomograph (Symphony 1.5 Tesla, Siemens) with a T1-weighted magnetisation prepared rapid gradient echo (MPRAGE) sequence with 1.3 mm slice thickness, 0.5 mm squared pixel size and 512 × 512 matrix. For testing imaging accuracy, additional scans with standard navigation protocols were obtained: a Picker Edge 1.5 T MR tomograph (T1 weighted flash-3D sequence, 135 slices, 1.3 mm slice thickness, 256 × 256 matrix, 1.0 mm squared pixel size), a Siemens Magnetom Open 0.2 T MR tomograph (T1-weighted flash 3D, 96 slices, TE 12.0 ms, TR 34 ms, field of view 250 × 250, 2.0 mm slice thickness, 256 × 256 matrix) and a Siemens Trio 3 T MR tomograph (T1 weighted MPRAGE, 1.3 mm slice thickness, 512 × 512 matrix, 0.5 mm squared pixel size) .

Definition of accuracy terms

  1. (a)

    Software accuracy: coordinates measured on image data with the system software compared to coordinates measured with a coordinate high precision machine [computer numerical control (CNC)] (Fig. 4).

    Fig. 4
    figure 4

    Flow chart showing types of measurements (image data set coordinates, computer numerical control coordinates and navigation coordinates), comparisons and transformation as well as types of accuracy (software, system and navigation accuracy)

    Full size image
  2. (b)

    Imaging accuracy: software accuracy subclassified with respect to the imaging modalities.

  3. (c)

    System (or digitising-) accuracy: navigation coordinates compared to CNC coordinates. It represents the accuracy of system components with which the digitising unit achieves calculation of a point’s space coordinates and includes precision of stereoscopic cameras and pointer instruments.

  4. (d)

    Navigation accuracy: coordinates from the navigation procedure compared to coordinates on image data sets. It represents the accuracy of the localisation of target points after registration and represents the practical overall accuracy, as it includes the digitising component, the software and imaging modality, the registration procedure as well as various intraoperative factors. Including all these possible error sources, a larger deviation than for the other types of accuracy was expected.

Types of measurements conducted

  1. (a)

    Computer numerical control machines, representing high precision standard, were used to obtain true coordinates of target points as a reference set. The phantom was measured (Fig. 5) with an accuracy of 5 μm (Global Image 9158, Brown & Sharpe, SN: 00072, QS Engineering, Bad Friedrichshall, Germany).

    Fig. 5
    figure 5

    The new phantom and its true target coordinates (x, y, z coordinates in millimeters) as measured by a Computer Numerical Control high precision machine: the main view (a) over the skull model and the small lateral view (b) show the deformable layer, adhesive fiducials (rings), two external target points (yellow dots) and openings towards the internal targets (red dots)

    Full size image
  2. (b)

    Target point coordinates, localised with the software of each system on image data sets (ten repetitions).

  3. (c)

    Navigation coordinates were measured with both systems in a clinical setup (Figs. 1 and 2). With Stryker 6 to 9, fiducials in an asymmetric non-coplanar pattern surrounding the lesion were used for registration [34, 39]. The camera was warmed up and positioning was checked, so that the phantom, patient tracker and pointer were well inside the camera field in a nearly ideal distance (1.5–1.8 m). After registration, the RMS was recorded; RMS more than 2.5 mm was not accepted. Success of every registration was controlled visually by pointing at well-defined structures. Each target was touched with the pointer tip, and with the “freeze” function, coordinates were acquired. For each measurement series, the procedure was repeated ten times to avoid systematic error factors, and each coordinate measurement was repeated three times to avoid random errors when pointing a target point, which was taken into account during statistical processing. To examine factors such as replacement of non-sterile instruments with sterile ones after registration or the use of a sterile drape over the patient tracker to avoid replacement, extra measurement series were performed.

    Slight modifications were necessary with BrainLAB due to the user interface. Target points were defined as “labelled points” and uniquely numbered. Patient data was loaded, and a new standard registration began with the pointer. Seven fiducials were used, which is the maximum number accepted by the system. An asymmetric non-coplanar wide placed pattern was obtained. After registration and visual checking of accuracy at well-known structures, such as the nasion, all target points were touched with the pointer and acquired as “intraoperative points,” as the system does not directly display coordinates. The “pointer to target” distance in millimetres was recorded for every intraoperative point and its closest labelled point, which represents the deviation between image data set coordinates and navigation coordinates. After each series three log files were copied, which contained all necessary information, including coordinates of target points and acquired points as well as a precision estimation of the registration. The same strategy for repeated measurements was followed.

  4. (d)

    Surface matching measurements were conducted with both systems. With Stryker these were carried out in analogy to the accuracy measurements. Two identical CT image data sets and two identical Symphony MRI data sets were used, differing only in the fixation during image acquisition. A control measurement of all target points with fiducial registration was conducted in order to obtain a reference. At least 30 surface points were acquired. The system suggests the acquisition of points from well-defined structures, such as the nose and the periorbital areas. Numbers of acquired and accepted surface points as well as “mean distance” were documented. Three head positions were examined: a normal supine position with asymmetrically widely distributed markers, a lateral position with almost linear placement of markers and a park-bench right position with markers occipitally, temporally and frontally. Registration of surface points on the nose or not was varied as a parameter. For each fiducial registration, all target points were measured with three to six consecutive surface registrations, while varying the examined parameters such as position, marked structures. etc. Fiducial only and fiducial plus surface registration were connected variables, which were taken into account for the statistical evaluation.

    Surface registration measurements with BrainLAB were conducted analogous to the corresponding accuracy measurements. Due to system limitations, the surface registration was not performed on top of the fiducial registration, but instead of it. The same four image data sets as with Stryker were imported. As being independent of previous fiducial registrations, double measurement series were necessary, i.e. extra fiducial registration as control. Surface points were acquired with a laser instrument (z-touch) plus an additional series with the mechanical pointer as a control. In addition, a series for testing the influence of different head positions and the corresponding acquired surface points was conducted. For this purpose, z-touch and pointer were combined, as in clinical routine. After registration, all target points were denoted with the pointer, and their coordinates were acquired as intraoperative points.

    For each measurement series, the procedure was repeated ten times, and each coordinate measurement was repeated three times to avoid random errors when pointing a target point.

Data transformation and statistics

Coordinate systems of data sets had to be matched, to permit a mathematical comparison. This was achieved through computed matching with the help of a well established fitting algorithm (Bevington’s “least squares fit”) [5, 48]. Residuals were calculated with a corresponding algorithm in R-project (R version 2.6.1). Transformation was only necessary for comparing CNC and navigation coordinates and CNC and image data coordinates. For the comparison of navigation with image data coordinates, no transformation was necessary, since both refer to the same coordinate system. Means and standard deviations were calculated as measures of accuracy and variance. Mixed effects regression models were used to account for the repeated measures and the hierarchical structure of the complex study design. These were used to examine the influence of parameters of the experimental setting on overall navigation accuracy as well as on software and system accuracy. Nested mixed effects regression models were compared to each other with the likelihood ratio test. All P values in this work refer to the comparison of nested regression models, unless explicitly mentioned. Due to the exploratory nature of the study, no adjustment was made for multiple testing. All tests were performed at a significance level of 0.05. Test results could not be interpreted in a confirmatory sense. Statistical analysis was performed using R-project (R version 2.6.1).

Results

A mean deviation for Stryker’s software of 0.38 ± 0.24 mm was calculated. Respective values for BrainLAB’s software were 0.44 ± 0.28 mm (Table 1). The difference was conspicuous (P < 0.001).

Table 1 Software, digitising (or system) and overall navigation accuracy (in italics) of the tested systems comparatively
Full size table

Examination of image data set accuracy showed a difference between CT and MR (Table 2). CT was more accurate with a mean Euclidean deviation of 0.28 ± 0.11 mm. All MR sequences were less accurate than CT (all P < 0.01). The worst accuracy was obtained with Open-MR (0.2 T intraoperative MR system). Corresponding analysis with BrainLAB showed similar results (Table 2).

Table 2 Mean deviation in millimeters ± standard error and 95% confidence intervals (CI) of image data examined with the software from Stryker–Leibinger and the respective software from BrainLAB
Full size table

Regarding accuracy of the digitising component, data showed a difference in favour of BrainLAB (Table 1). A mean Euclidean deviation of 0.72 ± 0.38 mm was calculated for Stryker as opposed to 0.33 ± 0.19 mm for BrainLAB (P < 0.01).

A larger deviation was expected for navigation accuracy, as it represents the overall application accuracy, including all possible error sources. This expectation was confirmed experimentally, yielding a mean Euclidean deviation of 1.45 ± 0.63 mm for Stryker and of 1.27 ± 0.53 mm for BrainLAB (Table 1). The difference of 0.18 mm was conspicuous (95% confidence interval: 0.09; 0.27, P < 0.01) but not constant in all three directions in favour of BrainLAB. Maximal outliers for both systems in all directions were of comparable magnitude.

Surface matching series resulted in deviations of 2.51 ± 1.49 mm for Stryker and 2.61 ± 1.56 mm for BrainLAB; these data show no significant difference (Table 1). Markedly larger deviations were observed from outliers for both systems with surface matching.

Influencing factors on fiducial registration

  • No difference was discovered between Stryker’s long or short, sterile or non-sterile pointer. Nevertheless, one of the non-sterile long pointers worsened accuracy by 0.55 mm (95% confidence interval: 0.10; 0.99; P = 0.02). Although the specific pointer had a visible geometrical deformation, the validation procedure was successful. In Stryker, touching a predefined point on the tracker with the pointer tip performs this. In BrainLAB, such a procedure is replaced by the position of the pointer in the instrument case, where the two pins in predefined positions allow a visual check of intact geometry.

  • Replacement of non-sterile instruments after registration with Stryker proved to decrease system accuracy by 0.11 mm marginally (95% confidence interval, 0.01; 0.20, P = 0.03) and navigation accuracy conspicuous by 1.13 mm (95% confidence interval, 0.52; 1.76, P < 0.01). This is possibly due to relative movement of the patient tracker in relation to the patient’s head during replacement. An alternative is the use of a sterile drape to cover the tracker. This method also increased the observed deviations as compared to no sterile drape. System accuracy was worsened by 0.17 mm (95% confidence interval, 0.11; 0.23, P < 0.01) and navigation accuracy by 0.69 mm (95% confidence interval: 0.28; 1.11, P < 0.01).

  • Results from regression analysis showed no relation between precision data given by the two systems (RMS and worst fiducial deviation displayed by Stryker, log-file data in BrainLAB) and digitising or navigation accuracy. The “pointer to target” distance calculated by the BrainLAB system, however, was related with navigation accuracy (P < 0.01).

Influencing factors on surface matching

  • A conspicuous difference for both systems in favour of CT, regarding software accuracy. The difference was 0.25 mm (95% confidence interval, 0.21; 0.28, P < 0.01) for Stryker and 0.28 mm (95% confidence interval, 0.26; 0.30, P < 0.01) for BrainLAB. In the Stryker system, this difference did not affect system- or navigation accuracy, i.e. imaging modality had no influence on the overall navigation accuracy. Nevertheless, the data show a difference for BrainLAB, where MR worsened the navigation accuracy by 0.71 mm (95% confidence interval, 0.28; 1.14, P < 0.01).

  • After surface matching, based on an existing fiducial registration (with Stryker), the data show no differences between the three tested head positions. In a park-bench right position of phantom, only with fiducial registration, a conspicuous difference was found, decreasing navigation accuracy by 1.67 mm (95% confidence interval, 1.39; 1.95, P < 0.01). This effect disappeared after the additional surface match, thus proving helpful in improving an inaccurate fiducial registration in this setting.

  • With Stryker, no influence on the mean deviation after surface matching was found regarding localisation of registered surface points or use of well-defined structures (e.g. nose), despite the subjective impression during measurements, that acquisition of points periorbitally and on the nose improved registration [28]. Acquisition of points on the nose as a separate factor could not be evaluated with BrainLAB, as they were almost always necessary for a successful registration; in such cases, a message would be shown from the system, requiring points from well-defined anatomical structures. Registration of periorbital points improved navigation accuracy with BrainLAB, decreasing the mean Euclidean deviation by 1.10 mm (95% confidence interval, −1.67; −0.53, P < 0.01).

  • The RMS and “mean distance” values from Stryker after combined fiducial and surface registration showed no influence on navigation accuracy, indicating that precision feedback does not always relate to true accuracy. Nevertheless, navigation accuracy was influenced by the number of acquired surface points (P = 0.01) and the percentage of accepted points for registration (P < 0.01), the product of both yields the total number of surface points actually being employed for registration. Feedback from BrainLAB, regarding achieved accuracy, can only be obtained retrospectively from a system log file. This value after surface registration proved a strong relation with navigation accuracy (P < 0.01).

  • Finally, a comparison between pointer, z-touch and combination of the two during registration revealed no differences regarding digitising or navigation accuracy.

Surface matching vs. fiducial registration

The digitising accuracy of both systems was marginally worsened with surface matching. Data showed conspicuous differences but were restricted in absolute value, thus undermining their clinical relevance (Table 3). A comparison of navigation accuracy with Stryker showed no difference between fiducial based and surface matching (P = 0.1910). The corresponding difference with BrainLAB (Table 3) was significant in favour of fiducial registration (P < 0.01). Mean Euclidean deviations of the two systems were of similar magnitude (Stryker 2.51 ± 1.49 mm; BrainLAB 2.61 ± 1.56 mm). Analysis of accuracy in spatial directions revealed a greater variance and more outliers in all directions with both systems after surface matching.

Table 3 Effect of surface matching compared to fiducial registration on digitising and navigation accuracy
Full size table

Discussion

In neuronavigation, knowledge of achieved accuracy and system limitations is crucial. Error analysis includes fiducial localisation error, deviation of fiducials after registration and deviation of points of interest other than the fiducials after registration [44]. Despite the fact that these errors represent vectors, they are given as values representing only their length, i.e. as “root mean square” (RMS) of the vector components. Since deviation of target points other than the fiducials [15], e.g. points in the vicinity of a deeply situated lesion, is hard to measure exactly with the navigation system, the surgeon must rely on a statistical estimation of this error, based on the known fiducial localisation accuracy or the error fit of the fiducials (RMS) [44]. Registration error calculated this way is nevertheless regarded by some authors as a non-trustworthy measure of registration accuracy [14, 41] and of real accuracy [39, 49]. After all, target registration error is considered anisotropic and may vary significantly in different areas of the head [43]. Different definitions [7, 25–27, 39, 44] and doubtful suitability of the systems’ feedback for the actual application accuracy make it necessary to apply methods of quantitative analysis with externally measured coordinates.

Quantitative feedback regarding registration precision given did not show any influence on accuracy, agreeing with previous studies [12, 39]. The RMS relies only on the geometric alignment of the fiducial markers and is not the accuracy that can be expected by the surgeon when approaching a target [1, 19, 46]. Disagreements between RMS and localisation accuracy are caused by the fact that the registration markers are only rarely evenly spread across the whole registration volume; only in this case—and if the object is spherical—the RMS may provide a good measure of the accuracy across the whole volume [12]. This raises the question how transparent systems are towards the surgeon, regarding the way they estimate achieved accuracy. Visualisation of quantified accuracy estimators on the image data in a 3D manner might prove useful in the future [32].

Imaging modalities

Advanced CT imaging accuracy was confirmed in this study, which could be explained due to MR distortion [24, 39, 50], being in accordance with previous studies [12, 34, 50]. However, no differences on overall navigation accuracy were detected. In the literature, conflicting results are reported in favour of CT [12, 34, 50] or MRI [18, 37]. The intraoperative MR system presented the most inaccurate results, which was expected from a low field system with accordingly low signal/noise ratio.

Comparison of the systems

In this study, the term navigation accuracy represents global accuracy, including all possible error sources, while system accuracy coincides with global accuracy in analogue studies [34]. Standardised reporting of error assessment, including the methodology used, may assure comparability of different accuracy reports [46].

Conspicuous differences detected in favour of BrainLAB regarding system and navigation accuracy are of limited clinical relevance, due to the submillimetric absolute values. The magnitude of the differences may be statistically significant for some measures, but in the context of the greater error seen in a clinical setting, these differences appear relatively unimportant.

Accuracy results of this work for Stryker are comparable to literature. A direct comparison is not feasible due to different imaging protocols and accuracy definitions. Poggi et al. performed a phantom study with the Stryker–Leibinger system [34]. They defined as global accuracy the difference in a reference set of coordinates acquired with a coordinate high precision machine with pointer coordinates after transformation. Influence of mechanical accuracy was not taken into consideration. Gellrich et al. [17] reported an overall accuracy of 1 mm for Stryker, but this was only the registration error reported and not the total experimental accuracy.

Regarding BrainLAB, results of this work are in accordance with previous studies. Values of 1.45 ± 0.99 mm are reported for phantom studies and 4.05 ± 3.62 mm for clinical studies [31]. Reports of registration error vary between 1.6 mm [8] and 2.1 mm [40]. Gumprecht et al. gave a target localisation error of 4.0 ± 1.4 mm, but it was an examination of a previous model with different methodology [20]. In vivo accuracy below 2 mm has been reported for the retrosigmoid approach using anatomical landmarks [9].

Influencing factors

Intact instrument geometry proved a necessity for accuracy. Validating the Stryker pointer with the patient tracker could sometimes be successful even with a deformed geometry, which could prove dangerous in clinical practice. This raises the question of whether validation thresholds should be re-evaluated, in order to find a balance between safe, strict validation on the one side and time-consuming validation failures and repeats on the other side.

Instrument replacement after non-sterile registration decreased accuracy conspicuously with Stryker, due to the change in relative position of the patient tracker in relation to the patient head (phantom) by the replacing procedure. The corresponding arm is robust, well fixed and has a specific interface for the patient tracker; nevertheless, a slight deviation cannot always be avoided. The alternative of covering the tracker with a sterile drape decreased accuracy as well. Refraction of infrared rays from the LEDs through the drape could be the reason. In passive systems like BrainLAB, problems are reported from the interruption of the visual line by the microscope drape [20].

A “park-bench right” position tested decreased navigation accuracy conspicuously. This effect disappeared after additional surface matching, indicating that surface registration may enable the system in this case to improve an inaccurate fiducial registration.

Fiducial vs. surface registration

Comparison of fiducial vs. surface registration showed a small difference in system accuracy of both systems in favour of fiducials, as well as a greater variance and more outliers with surface registration in all spatial directions. In general, algorithms for surface matching represent an effort to limit the influence of movable skin; nevertheless, they may lead to even greater registration errors [18, 22, 41], possibly due to sampling errors owing to uneven grouping of scalp points chosen for registration [18].

In daily clinical practice, registration with anatomic landmarks and/or surface matching is more practical. Insignificant loss of registration accuracy using natural anatomic landmarks compared to skin adhesive fiducials has been reported [50]. Nevertheless, fiducial registration alone remains the gold standard [22] of all non-invasive methods [33], being more accurate than the combination of anatomic landmarks and surface registration [18, 38]. Surface matching may produce larger deviations when added to a fiducial registration, and though it may improve accuracy after landmark registration, this combination fails to reach the level of fiducials alone [22]. In other studies, all three registration techniques (fiducials, anatomic landmarks and combined anatomic landmarks and surface matching) provided comparable deviations [49]. Mascott et al. reported of the superiority of implanted fiducials, but no major differences between adhesive fiducials, surface matching or anatomic landmark registration [28]. In recent studies, distribution templates for fiducials are provided, and a configuration with eight fiducials optimised over the whole head is proposed [42, 43].

Phantom critic

The deformable outer layer simulation provides an additional error source, playing a role not only during surface matching but also due to fiducial displacement. Skin displacements can cause problems during imaging, head fixation, sampling of fiducial markers or surface points for registration with the pointer [22, 38, 50]. Furthermore, rigid phantoms cannot simulate the contribution of soft tissue properties to MR errors [39]. The anthropomorphic form of the phantom allows the study of accuracy in relation to the position of target lesions, since larger deviations are often reported for posterior localisations [22, 36]. The disadvantage of non-anthropomorphic phantoms is the lack of correct anatomical simulation [4, 16, 23, 29, 30, 34]. Although the target volume may be similar to the real anatomical target volume, a substantial difference is seen in the registration process [46]. Anthropomorphic skull phantoms represent the bony anatomy, but it is very difficult to correctly simulate soft tissues [1, 12, 13, 21, 35]. Furthermore, the fact that the real head is not spherical and has a moveable scalp layer gives rise to areas with a real error substantially deviating from the RMS [12]. Skin flexibility and elasticity have been connected to problems in surface algorithms; pressure deformity during image acquisition and lack of skin turgor with presence of wrinkles in adults may all contribute to decreased accuracy using surface-fit registration [41].

The present phantom represents a novel effort towards realistic simulation of human-like anatomy and possible inherent inaccuracy factors. Such models are necessary for conducting high-quality phantom studies, which are essential to assure patient safety before clinical application, as well as for standardised evaluation and quality assurance of a constantly evolving technology [46]. Naturally, patient studies still remain essential, as it would be misleading to extrapolate phantom accuracy to true intraoperative conditions.

Conclusions

Experimental study of two optical systems provided evidence for the same level of accuracy. Passive frameworks need not necessarily be less accurate than active LED systems. Overall navigation accuracy of both systems was definitely acceptable (< 1.5 mm), while surface registration failed to show better results than fiducial registration. The newly developed phantom allows the experimental study of surface registration and represents a novelty towards non-rigid phantoms.

Accuracy feedback by the navigation systems should be considered with caution, as it does not always correlate well with real localisation accuracy and is therefore to be seen only as an averaged estimation for the quality of the registration procedure. Checking accuracy visually directly after registration with the help of anatomical landmarks on the surface ought to be repeated during the operation by verifying with well-defined deeper structures.

Distinguishing different types of accuracy experimentally may allow identification of error sources, and consideration of all potential influencing factors may minimise errors in navigation procedures.