Probe versus microscope: a comparison of different methods for image-to-patient registration

Clinical navigation systems are commonly used for ENT interventions. At critical regions, they are supporting the surgeon by realizing the exact positions of a pointer inside the patients’ body on the CT/MR images. Depending on the surgical intervention, different pointers can be used for navigation; for example, a microscope is required at the lateral skull base to support the surgeon with detailed images of the surgical area. In this case, the pointer can be the microscope or a probe.

Prior to using a navigation system, the CT/MR images have to be registered to the patient. In general, the two main registration methods are paired-point and surface registration [1, 2]. For the first method, points on the patient (‘fiducials’) have to be located, together with the corresponding points in the images. Thus, the rigid transformation between image and patient coordinates can be calculated. Besides, anatomical structures also skin-affixed or bone-implanted markers can be used as fiducials. Skin-affixed markers are known to be prone to changes of the skin according to different hydration statuses or effects of muscle relaxing agents and thus are to be used with caution [3]. Those aspects can be avoided if bone-implanted fiducials are used, but because of the invasiveness they are not the first choice, due to side effects for the patient. The most commonly used fiducials are anatomical landmarks, as they are easy to use and non-invasive.

Surface registration helps automation and registers a larger number of points to the image. These points are collected by moving a pointer over a large surface on the patient, like the forehead. With the iterative-closest-point (ICP) algorithm, the surface of the patient is matched to the corresponding 3D-image-model [2].

For interventions, e.g. the lateral skull base or the petrous bone, the VBH (Vogele–Bale–Hohner) head holder can be used additionally to fix the patients head on the OR Table [4]. External fiducials can be affixed to the VBH mouthpiece to get a larger area of fiducials that surrounds the surgical region and thus improve accuracy [5, 6]. More details are given in the ‘Materials and methods’ section.

The ideal case for the surgeon would of course be perfect (error free) navigation, but the registration is a critical step and has to be done as accurately as possible. However, it always introduces errors as calibration, tracking and measurement errors can occur, but user localization errors are also common in surgical navigation [7‐10].

After registration, the typical navigation systems report a single error which is called the fiducial registration error (FRE) [11]. The FRE is the RMS distance of CT image fiducials and patient fiducials that are transformed to the CT image. However, the target registration error (TRE) is much more important for the surgeon in practice. It reports the accuracy at a target location close to the surgically relevant positions and is uncorrelated with the FRE [12].

Unfortunately, not many studies regarding the application accuracy (the TRE) of different registration methods have been done comparing microscope and probe registration methods [13‐19].

Therefore, the aim of this study was to find the optimal registration method of a plastic skull to its preoperative CT images in terms of application accuracy at the lateral skull base, concerning the pointer and the registration method. Two different questions were evaluated: What kind of pointer should be used for registration and measurements? Which registration method is the optimal method to register a plastic skull to its CT images?

Two different pointers were used for both registration and TRE measurement: a microscope and a pointer. Both pointers are optically tracked by a camera relative to a DRF (digital reference frame), so the coordinates of the points in patient space can be determined and transformed to image space.

Four different registration methods have been experimentally evaluated with each pointer (probe or microscope), two of them with external fiducials on the VBH mouthpiece. Surface registration was additionally used for half of the experiments (see Table 1 for a detailed description). The TRE was measured on 10 different targets, and all methods were compared to each other. Thus, it was possible to evaluate the performance and differences of the registration methods and pointers used for our experiments.

The experimental set-up can be seen in Figs. 1, 2.

All results are mean values with their standard deviation and can be seen in detail in Tables 2, 3. Group A defines the registration with the probe and group B the registration with the microscope.

All errors in the groups, except A1, A2 and B2, are normally distributed (\({p} < 0.05\)). All 8 groups compared to each other are not equal (ANOVA, \({p} < 0.0001\)); an overall difference of microscope- and probe-based registration can also be observed (Wilcoxon signed rank, \({p} < 0.0001\)).

The overall TRE for group A is \(1.69 \pm 0.74 \, \hbox {mm}\), and for group B, it is \(2.19 \pm 0.94 \, \hbox {mm}\).

The mean TRE of A1 is 1.41 mm compared to B1 with a mean TRE of 1.70 mm. The mean TRE of A2 is 1.48 mm compared to B2 with a mean TRE of 3.02. A3 and B3 have a mean TRE of 1.44 mm and 1.92 mm, respectively. A4 and B4 have a mean TRE of 2.44 mm and 2.10 mm, respectively (Table 2).

These results are leading to statistical results as follows: A1 is equal to B1 (\(p = 0.0645\)) and A4 is equal to B4 (\(p = 0.1602\)). A2 and B2 are significantly different (\(p = 0.0020\)), as well as A3 and B3 (\(p = 0.0273\)).

The four probe groups (A1–A4) are significantly different to each other (\(p = 0.0004\)). A4 has the largest error.

B1–B4 are significantly different to each other (\(p = 0.0012\)), and B2 has the largest TRE.

Regarding the error in x-, y- and z-directions, it could be observed that overall the TRE is anisotropic. For the probe measurements, the error in y-direction is the smallest with \(0.79 \pm 0.49 \, \hbox {mm}\) and in x- and z-directions the error is \(1.00 \pm 0.56 \, \hbox {mm}\) and \(1.03 \pm 0.62 \, \hbox {mm}\), respectively. This is also valid for the microscope; in y-direction TRE is \(1.08\pm 0.60 \, \hbox {mm}\), in x-direction \(1.31 \pm 0.80 \, \hbox {mm}\) and in z direction \(1.27 \pm 0.72\,\hbox {mm}\).

The TRE of anatomical targets and skin-affixed markers can be seen in Table 3. Skin-affixed markers have a TRE of \(1.50 \pm 0.67 \, \hbox {mm}\) for group A and \(2.16 \pm 1.02 \, \hbox {mm}\) for group B. The TRE for the anatomical landmarks is \(1.82 \pm 0.77 \, \hbox {mm}\) (group A) and \(2.10 \pm 0.78 \, \hbox {mm}\) (group B).

Dividing skin-affixed markers in radiolucent and donut markers, the TRE of group A is \(1.16 \pm 0.40\) and \(1.84 \pm 0.85 \, \hbox {mm}\), respectively, and of group B \(1.62 \pm 0.71\,\hbox {mm}\) and \(2.71 \pm 1.27\,\hbox {mm}\), respectively.

Pointwise evaluation is shown in Fig. 7. Target number 10 (right infraorbital foramen) is the only point, where using the probe leads to a larger TRE than using the microscope.

The results of the worst-case scenario can be seen in Tables 4, 5. The TRE for group A is \(1.98 \pm 0.80\,\hbox {mm}\) which means a deterioration of the mean of 0.29 mm in comparison with the overall result. For group B, the worst-case TRE is \(2.49 \pm 0.99 \, \hbox {mm}\); this is a deterioration of the mean of 0.30 mm compared to the overall TRE.

The worst-case TRE for group A and skin-affixed markers is \(1.77 \pm 0.71\,\hbox {mm}\) (\(1.29 \pm 0.41\,\hbox {mm}\) radiolucent markers and \(2.26 \pm 0.92 \, \hbox {mm}\) donut markers) and \(2.10 \pm 0.83 \, \hbox {mm}\) for anatomical landmarks (Table 5).

Group B has a worst-case TRE of \(2.41 \pm 0.81\,\hbox {mm}\) for the anatomical landmarks and \(2.42 \pm 1.09\,\hbox {mm}\) (\(1.75 \pm 0.73\) radiolucent markers and \(3.10 \pm 1.35\,\hbox {mm}\) donut markers).

This study evaluated four different registration methods using two different pointers. From a first point of view, it was apparent that the microscope led to larger TREs than the probe. However, with additional surface registration plus using the VBH mouthpiece, the probe registration led to larger TREs than the microscope. Statistically no difference between microscope and probe registration could be found in this case. In general, more errors are influencing the accuracy of the microscope compared to the probe. Both pointers are tracked, but the error chain of the microscope also includes camera calibration transformation errors. From this aspect, it can also be argued that the accuracy of the microscope is similar to that of the probe.

In general, surface registration is more time intense than paired-point registration. But also locating the fiducials and targets with the microscope takes much more time due to difficulties in focusing the points. Those two facts were already speaking against using the microscope and an additional surface registration before any calculations were done. However, it is evident that some interventions, e.g. at the petrous bone, cannot be done without a navigated microscope and thus the time factor is negligible.

The worst results can be observed for B2 and A4, i.e. paired-point combined with surface registration is less accurate/precise than paired-point registration only. This result was also found in [14].

However, there was no (statistically significant) difference between A1, A2 and A3. All those methods led to a mean TRE lower than 1.5 mm. Though it has to be said that A1 was the method that needed the fewest preparations and was ‘easy to use’. Thus, paired-point registration only is an adequate method for our experiments.

Regarding the registration with the microscope, B1, B3 and B4 were statistically equal. Using the microscope was more time-consuming due to the waiting time until the microscope is at rest and the autofocus is adjusted to confirm a fiducial/target. However, when only paired-point registration was applied, the time factor can be neglected and the TRE was similar to the experiments done with the probe.

Pointwise evaluation clearly showed that donut markers were the worst targets to use. This may be caused by the form of this marker, which is difficult to focus with the microscope. It is also possible that those markers were difficult to locate in the CT images, or the centre might not have been clearly visible. For our experiments, it would have been better not to use donut markers as a target. Anatomical landmarks showed better TREs, the smallest values could be obtained at X-spots. The targets with the largest mean TRE when using the probe were the anatomical marker target nr. 9, and the Donut marker target nr. 5 when the microscope was used.

It is clear that on a plastic skull anatomical landmarks and skin affixed markers can be located more accurately in contrast to a real patient, where the skin is moving, or bony structures, like foramina cannot be touched precisely. Besides that, it is hard to realize not using anatomical landmarks, because they are easy to use, non-invasive and clinical standard.

Anisotropy of the TRE in x-, y- and z-directions could be detected. In our experiments, the mean TRE is about 0.3 mm smaller in y-direction than in x- and z-directions. Typically, in optical systems the viewing direction has the largest noise variance, e.g. [20]. The anisotropy found in our set-up can be caused by not orienting the probe/microscope correctly to the camera, the underlying FLE, or the variance of the camera. Besides that, the slice thickness of the CT images is larger than the pixel resolution of the images.

The effect of the location of the fiducials that are used for registration is an interesting point which can influence the accuracy of navigation. The fiducials for paired-point registration are the same for A1/B1 and A2/B2, but also for A3/B3 and A4/B4. As shown in the results only A4 and B2 are significantly different to the other groups. But both methods are using surface registration in addition that should improve results, especially as a rigid surface was used [21]. The standard deviations for both A4 and B2 are about the double of the standard deviations of A1, A2, A3 and B1, B3, B4, respectively. Registration, measurement or user errors could be the cause for this discrepancy. Besides that, the difference seems to be caused due to difficulties in locating the targets. Thus, it seems that external fiducials on the VBH mouthpiece do influence the accuracy of navigation in a negative way when a microscope is used.

Analysing the worst-case scenario, it can be observed that the deterioration of the mean TRE is in all cases not more than 0.35 mm, the standard deviation staying quite constant for all cases. This means that measurements were good and constant, without large outliers. Considering the different types of targets, the values of the X-spots increase about 0.1 mm, but Donut markers have mean values more than 0.5 mm worse. Anatomical landmarks have a mean TRE that is about 0.4 mm worse than in the overall case. This speaks again against using Donut markers but also against anatomical landmarks, because locating them is difficult and leads to a large error. In contrary to the overall scenario, target 5 has the largest mean TRE for both groups A and B in the worst-case scenario. This target is a Donut marker on the left parietal bone.

Overall the largest TREs can be observed on the targets 5 and 9 for both groups (Donut marker on the left parietal bone and the saddle point on the right frontozygomatic suture), the smallest TREs on targets 1 and 2 (X-spots). Due to the experimental setting, locating target 5 might be difficult (Fig. 8). Target 9 is a very unspecific point and is difficult to locate both in the CT image and on the plastic skull, in contrary to targets 1 and 2, which are X-spots, that are easy to locate on the front of the skull. X-spots can be located more precisely than all other markers, they have the smallest standard deviations. In contrary to that, Donut markers have the largest standard deviations; only in experiments A4 and B4, anatomical landmarks can be located less precisely.

Unfortunately, the study is limited to a paired-point registration using 10 fiducials throughout the experiments. As already confirmed by other studies (e.g. see [8, 10, 22]), a smaller number of fiducials lead to a larger TRE than using more fiducials for registration. However, using, for example, 5 fiducials is not necessarily leading to a larger TRE than using 10 fiducials. It is always depending on the situation and on the fiducials used. Due to not using a different number of fiducials for the same experiment, we do not know if this would lead to an improvement for one of the registration methods. However, the 10 fiducials used were placed ideally on the plastic skull to ensure a good configuration for the registration process.

The target registration error of different registration pointers and methods was evaluated, and differences between the alternative approaches could be observed. Overall it can be observed that a probe used with paired-point registration gives the best results for navigated surgery. Additional surface registration and/or fiducials do not improve the TRE significantly in our experiments. Using a microscope as a pointer leads to a larger error compared to the probe for all experiments except when paired-point registration is combined with external fiducials and surface registration. However, in clinical practice pointers are mostly used for registering the patient to pre- and intraoperative radiological data.

In conclusion, both registration instruments had comparable application accuracies for the plastic skull in an experimental set-up. Using the microscope is a rarely used registration device, but this study demonstrates that registration outcome is not affected by the registration tool chosen.