2.1 Patient cohort and image acquisition
For this study, ten head and neck (HN), ten gynecological (GYN) and ten lung (LNG) cancer patients treated with IGRT using CBCT were retrospectively selected. The HN and GYN patients received volumetric modulated arc treatment (VMAT) using either 6 MV or 10 MV photon beams with doses up to 70 Gy in 35 fractions (HN) and 45 Gy in 25 fractions (5 fractions/week) prescribed at the median dose of the high risk PTV (GYN), respectively. LNG patients received stereotactic body radiation therapy using 3D-conformal radiotherapy using 10 MV FFF beams with doses of 60 Gy in eight fractions or 45 Gy in three fractions prescribed to the 65% isodose covering 99% of the PTV. The planning CTs (pCT) of the patients were acquired with a Somatom AS (Siemens AG, Forchheim, Germany) CT using a slice thickness of 2 to 4 mm depending on the treatment. All images were reconstructed using iMAR (Iterative metal artifact reduction) if necessary due to implants or dental fillings. The patients were treated using VersaHD linear accelerators (Elekta AB, Stockholm, Sweden). Consequently, CBCT image data was acquired using the XVI kV imaging system (Elekta, Stockholm, Sweden). The standard protocols provided by Elekta were used in all cases, i.e. ‘Fast Head and Neck S20’, ‘Pelvis M20’, ‘Chest M20’ presets were used for the acquisition of the CBCT data for the HN, GYN and LNG patients, respectively. The CBCTs were using the same slice thickness as the pCT.
2.2 Algorithms
All deformations and dose calculations were done in a research version of the RayStation TPS (V. 11B-DTK, RaySearch Laboratories, Stockholm, Sweden) and dose calculation was performed using its collapsed cone algorithm.
The novel CBCT correction algorithm is similar to an algorithm proposed by Shi et al. which also uses deformable image registration of the pCT to the CBCT for the correction of image intensities and shading artefacts of the CBCT. In short, the algorithm calculates a correction map based on the differences between the CBCT and the pCT. This correction map is used to enhance the quality of the CBCT. Further, the algorithm analyses the image intensities of different tissue types in the CBCT and the pCT and calculates calibration curve using piecewise linear interpolation between these tissue types. Additionally, the algorithm employs a stitching technique to simulate missing tissue outside the field of view (FOV) of the CBCT by attaching the pCT outside the FOV [20,21,24]. Early versions of the algorithm using the scripting interface of the TPS have been tested have been tested by other groups [21–23]. The output of the algorithm is a so called corrected CBCT (CBCTc) which can be used for dose calculation using the CT to mass-density conversion curve of the pCT.
In addition to the novel algorithm, the standard algorithm for CBCT dose calculation implemented in the TPS was used for comparison. This algorithm uses a threshold-based bulk density overriding technique. Six different tissues-types can be segmented based on the image intensity. A standard density is assigned to these segmented tissue types. The thresholds of the individual tissue types need to be adjusted for each image to yield optimal calculation results. The output of this algorithm is a step-wise function converting CBCT image intensities to mass density to enable dose calculation. Dose distributions calculated using this method are labeled CBCTb.
2.3 Analysis
The workflow or the analysis of the data is shown in Figure 1. For the HN and GYN patients the pCT was rigidly and in a subsequent step deformably registered to the CBCT to generate a deformed CT (dCT) which has similar anatomical features as the CBCT to reduce the influence of anatomical changes between the pCT and the CBCT. The settings for the deformable image registration are shown in Table 1. The same deformable registration was used for generating the dCT and the CBCTc. The deformation field was visually inspected for unreasonable and particularly large deformations. The generation of the dCT was omitted for the LNG patients since large parts of the body were outside the FOV of the CBCT and therefore introduced large artefacts while deforming the CT. Thus, all LNG patients were only analyzed using the pCT as ground truth. The CBCTc and the CBCTb were generated for all patients using the respective algorithms. For each patient in the HN group, the original treatment plan was re-calculated using the pCT, dCT, CBCTC and the CBCTB within the region of the body fully covered by the FOV of the CBCT as “External” contour. For each patient in the GYN and LNG group, the original treatment plan was re-calculated using the pCT, dCT, CBCTc and the CBCTb with the FOV of the CBCT as “External” contour. This means that dose was calculated in the same region for each image set and therefore allows a fair comparison of both CBCT conversion methods. No further optimization of the treatment plan was performed.
Table 1 Parameters used for the deformably image registration between pCT and CBCT.
Parameter
|
Value
|
Deformation strategy
|
Default
|
Similarity measure
|
Correlation coefficient
|
Deformation grid resolution
|
0.25 cm x 0.25 cm x 0.25 cm
|
Focus ROI
|
CBCT FOV contracted by 2 cm
|
To assess the performance of the stitching technique, the dose calculated on the pCT and the CBCTc was recalculated for the HN and GYN cases using the original body contour as external contour. The LNG cases were excluded from this evaluation as the beam entrance and the target volume were not affected by the FOV of the CBCT.
All dose files were exported and analyzed using the MICE toolkit (NONPIMedical AB Sweden, Umeå). Gamma analysis was performed for four different dose threshold levels: 10%, 30%, 50% and 90%, relative to the prescribed dose, performing local gamma analysis with a 1% dose difference and a 1 mm distance to agreement criterion. The resulting gamma pass rate (GPR) defined as the percentage of gamma indices smaller than or equal 1 was recorded.
The differences between the investigated groups were analyzed statistically using the Wilcoxon rank sum test considering a p-value smaller than 0.05 as statistically significant. Statistical computing was performed using R (V. 4.2.2, R Foundation for Statistical Computing, Vienna, Austria).