The spatial accuracy of two frameless, linear accelerator‐based systems for single‐isocenter, multitarget cranial radiosurgery

Abstract Single‐isocenter, multitarget cranial stereotactic radiosurgery (SRS) is more efficient than using an isocenter for each target, but spatial positioning uncertainties can be magnified at locations away from the isocenter. This study reports on the spatial accuracy of two frameless, linac‐based SRS systems for multitarget, single‐isocenter SRS as a function of distance from the isocenter. One system uses the ExacTrac platform for image guidance and the other localizes with cone beam computed tomography (CBCT). For each platform, a phantom with 12 target BBs distributed up to 13.8 cm from the isocenter was aligned starting from five different initial offsets and then imaged with the treatment beam at seven different gantry and couch angles. The distribution of the resulting positioning errors demonstrated the value of adding a 1‐mm PTV margin for targets up to about 7–8 cm from the isocenter. For distances 10 cm or more, the CBCT‐based alignment remained within 1.1 mm while the ExacTrac‐based alignment differed by up to 2.2 mm.

2 | ME TH ODS AND MATERIALS

2.A | Phantom design
When this study was initiated, there was no commercially available head phantom with multiple radio-opaque targets distributed throughout the bony structure of a skull. We therefore constructed a phantom from readily available materials: three sections of wood beam of nominal commercial cross-section 4″ 9 4″ (actual dimensions approximately 8.7 9 8.7 cm 2 ) were cut to lengths of approximately 20 cm and glued together after first embedding in them 12 chrome steel ball bearings of diameter 4.8 mm (3/16″). Figure 1 shows a photograph of the phantom and orthogonal radiographs demonstrating the distribution of the targets. The most central target was designed to be at the isocenter for the subsequent plans. The radial distance of the targets from the isocenter ranged from 3.1 to 13.8 cm. This range was chosen because in some cases one might choose to put the isocenter on a target near a critical structure, such as the brainstem or the optic chiasm, instead of in the center of the distribution of targets.
Although the internal structure of the phantom is not anthropomorphic, the wood grain permeates the phantom and is used in the image guidance process (Fig. 2). This has some advantages over phantoms that embed BBs in plastic cubes, because in this phantom, while the target BBs appear in the images, they contribute relatively little to the overall information used to drive the image guidance. This is more like the clinical situation in which one aligns to the bony anatomy and cannot visualize the actual targets with the imaging used for alignment. via an optical surface-monitoring system, but that system was not employed for this study. The phantom was initially aligned using CBCT and no adjustment was made after each couch rotation. The HU of the wood grain varied between approximately À800

2.C | Treatment plan
and À250. While that structure showed up clearly on the planning CT, CBCT, and Eclipse DRRs, it entirely disappeared on the DRRs constructed by the ExacTrac system's algorithm during its alignment process. For that reason, a second CT study was artificially created using MATLAB (MathWorks, Natick, MA) code that increased the HU of the wood structure to the range of 50-800, mimicking soft tissue to bone.
This was the study used in the testing of the Varian iX with ExacTrac.  Table 1 shows the range of the initial corrections applied in the five iterations of each process.

2.E | Image analysis and offset measurements
Each of the acquired MV images was opened in the Aria Offline Review module (Varian Medical System, Palo Alto, CA, USA) as an overlay to its associated DRR. Using the distance measuring tool, the offset between the center of the target in the DRR and that in the MV image was measured and recorded. The 12 targets were measured for the seven fields for the five repetitions on the two linacs for a total of 840 measurements. One field was measured on three occasions to test the reproducibility of this manual process.
Each of the 12 targets was therefore imaged 35 times on each linac. The average offset between the treatment field and DRR was calculated along with the maximum and the standard deviation. The 95% confidence limit on the offset was estimated by adding the average with twice the standard deviation. These results were then plotted as a function of the distance of the target from the isocenter.
3 | RESULTS Table 2 shows the average offset, maximum offset, standard deviation and 95% confidence limit as a function of distance from the isocenter for the two linacs and image guidance processes.  Figure 4 shows the 95% confidence limits for the two processes on a single graph.
For the target planned to be at the isocenter, the 95% confi-  Table 2 range from 0.2 to 0.5 mm, so one can conclude that the uncertainty associated with this manual measurement technique contributed to but did not dominate the variability in the observations.

| DISCUSSION
In the study, the offsets seen for the ExacTrac system at 10 cm from the isocenter and beyond were more than those seen for the CBCTbased alignment system. This may be attributed to two factors. First, the ExacTrac tolerances that were applied, 0.7 mm and 0.7 degrees, inherently permit some variation that will be magnified at distance.
Second, the field of view of the ExacTrac images (13 9 13 cm 2 ) is  Another potential weakness of this nonanthropomorphic phantom is that the grain structure on the ExacTrac radiographs approximate parallel lines, as seen in Fig. 2. This is very different from the appearance provided by a human skull, so it is possible that the ExacTrac algorithms would be more accurate with a realistic phantom. It is hoped that phantom vendors will soon provide such tools so that physicists can perform end-to-end tests for multitarget SRS with realistic phantoms.
The observed offset between the target on the planned DRR and the MV image represents the geometric error in beam delivery that would have occurred for that target and field without regard for the direction of the offset. This study did not attempt to quantify the effect on dose coverage for a full treatment employing arcs at these couch angles, but it is intuitively clear that such errors will compromise target coverage. Other studies 9,10 have shown that the decrease in target coverage expressed as the dose covering 95% of the target (D95) is more significant for small targets.
T A B L E 2 Offsets between the planned and imaged target positions for the two platforms. For each target, the distance from the isocenter (in cm) and the average and maximum offsets between the planned and imaged targets (in mm) with the standard deviation in the 35 measurements for each target. The 95% confidence limit is approximated by the last column, which sums the average offset with twice the standard deviation. Stanhope et al. 9 analyzed sequential CBCT images for 22 patients who had SRS treatments to two targets with isocenters with the purpose of measuring the rotational difference in the skull between the two scans. By aligning the second scan to the first, they measured how much the patient had moved within the mask and therefore estimated the uncertainty associated with patient motion after an initial alignment, which they termed "intraoperational" uncertainty. They found that 0.1 mm/cm of target-isocenter separation would account for 95% if this uncertainty. This assumed that the initial correction completely removed any initial setup error.
Roper et al. 10 (Fig. 2) for the case that six degree of freedom corrections are made, and use the curve for 1 sigma, which they found to correspond to 96.4% of the treatments. Reading from the graph, at isocenter the error is approximately 0.8 mm, at 3 cm it is 1.0 mm, and at 7.5 cm is 1.5 mm for the immobilization system used in their study and 1.5 mm, 2.0 mm, and 3.1 mm for the ExacTrac systems described by van Santvoort et al. 13

| CONCLUSION
What practical recommendations follow from these results? Most publications about single-isocenter, multitarget treatments advocate centering the isocenter among the targets. This will generally keep the distance from any part of a target to the isocenter to about 8 cm. At such distances, it would be prudent to increase the PTV margin by 1 mm. If one chose to place the isocenter on a target near the brainstem or chiasm while also treating targets near the superior aspect of the brain, then the distance from isocenter would be larger and an additional margin might be advisable, depending on the size of the lesions and the eloquence of the surrounding tissue. Adding margin is not necessarily benign. Kirkpatrick et al. 14 in a randomized trial involving 49 patients with 80 metastases, found more radionecrosis in those patients having a 3-mm margin compared to a 1-mm margin.
A fundamental outcome of this study is that it confirms that either platform in our institution is suitable for single-isocenter, multitarget SRS.

ACKNOWLED GMENTS
Thomas Boltz, Ph.D., provided the MATLAB code used in this study.