2D antiscatter grid and scatter sampling based CBCT method for online dose calculations during CBCT guided radiation therapy of pelvis

Abstract

Background:

Online dose calculations before the delivery of radiation treatments have applications in dose delivery verification, online adaptation of treatment plans, and simulation-free treatment planning. While dose calculations by directly utilizing CBCT images are desired, dosimetric accuracy can be compromised due to relatively lower HU accuracy in CBCT images.

Purpose:

In this work, we propose a novel CBCT imaging pipeline to enhance the accuracy of CBCT-based dose calculations in the pelvis region. Our approach aims to improve the HU accuracy in CBCT images, thereby improving the overall accuracy of CBCT-based dose calculations prior to radiation treatment delivery.

Methods:

An in-house developed quantitative CBCT pipeline was implemented to address the CBCT raw data contamination problem. The pipeline combines algorithmic data correction strategies and 2D antiscatter grid-based scatter rejection to achieve high CT number accuracy. To evaluate the effect of the quantitative CBCT pipeline on CBCT-based dose calculations, phantoms mimicking pelvis anatomy were scanned using a linac-mounted CBCT system, and a gold standard multidetector CT used for treatment planning (pCT). A total of 20 intensity-modulated treatment plans were generated for 5 targets, using 6 and 10 MV flattening filter-free (FFF) beams, and utilizing small and large pelvis phantom images. For each treatment plan, four different dose calculations were performed using pCT images and three CBCT imaging configurations: quantitative CBCT, clinical CBCT protocol, and a high-performance 1D antiscatter grid (1D ASG). Subsequently, dosimetric accuracy was evaluated for both targets and organs at risk as a function of patient size, target location, beam energy, and CBCT imaging configuration.

Results:

When compared to the gold-standard pCT, dosimetric errors in quantitative CBCT-based dose calculations were not significant across all phantom sizes, beam energies, and treatment sites. The largest error observed was 0.6% among all dose volume histogram metrics and evaluated dose calculations. In contrast, dosimetric errors reached up to 7% and 97% in clinical CBCT and high-performance ASG CBCT-based treatment plans, respectively. The largest dosimetric errors were observed in bony targets in the large phantom treated with 6 MV beams. The trends of dosimetric errors in organs at risk were similar to those observed in the targets.

Conclusions:

The proposed quantitative CBCT pipeline has the potential to provide comparable dose calculation accuracy to the gold-standard planning CT in photon radiation therapy for the abdomen and pelvis. These robust dose calculations could eliminate the need for density overrides in CBCT images and enable direct utilization of CBCT images for dose delivery monitoring or online treatment plan adaptations before the delivery of radiation treatments.

1. Introduction

Over the past two decades, CBCT imaging has primarily played a role in radiotherapy by ensuring accurate target localization. This ensures that the prescribed dose is delivered to the targets while sparing surrounding normal tissues^1,2. Although this approach is effective for target localization, it does not allow for verification of dose delivery to the targets and surrounding normal tissues before or during radiation treatment. Factors such as weight loss, changes in tumor size, and spatial changes in normal tissues during treatment can lead to discrepancies between planned and delivered doses. To address this, the concept of image guidance has evolved from solely target localization tasks to include online dosimetric verifications using volumetric images acquired prior to treatment delivery^3–7.

However, a well-known challenge in utilizing CBCT images for radiotherapy dose calculations is the issue of dosimetric accuracy. The relatively poorer CT number accuracy in CBCT can hinder the precise extraction of mass or electron density from these images, which is essential for accurate dose calculations^8,9. Over the years, numerous solutions have been proposed to address this issue. One of the most commonly employed approaches is assigning density, or density override, to anatomical structures. This allows for dose calculations to be performed by utilizing pre-defined density information^8,10,11. Density overrides often necessitate the segmentation of anatomical structures, and the precise density of these structures is not known but rather estimated. To overcome the challenges associated with density overrides, alternative approaches such as anatomy and patient size-specific Hounsfield Unit (HU) to density tables have been proposed^12–14. An alternative approach involves fusing the planning CT and CBCT images using deformable registration methods. This allows for the transfer of HU values or dose information from the planning CT onto the CBCT images^6,15–20. However, correlation of anatomical regions and densities between planning CT and CBCT after deformable image registration may contain registration errors^19,21,22. More recently, Deep Learning methods have been investigated extensively to generate high quality CBCT images from standard CBCT images and learning from the gold standard pCT images, known as synthetic CT^23–27. While these methods can achieve remarkable CT number accuracy, their ability to consistently generate high-fidelity synthetic CT images for various patient sizes, anatomical regions, and radiotherapy treatment setups is still under investigation.

More accurate dose calculations can be achieved by physics-driven strategies via improving the CT number accuracy of CBCT images. Such strategies aim to mitigate raw CBCT data contamination. For example, scatter is one of the major reasons behind raw data contamination in CBCT. Improved scatter correction methods can improve CT number accuracy, and dose can be calculated more accurately by using the CBCT images directly^28–31. This approach also simplifies the clinical workflow and reduces potential dosimetric errors associated with density overrides, deformable image registration, and synthetic CT generation. However, one potential drawback of existing methods is the accuracy of the achieved HU values through these physics-driven approaches. Due to the diversity in patient sizes and treated anatomical regions, errors in HU accuracy can lead to dosimetric errors in CBCT-based dose calculations.

In this study, a novel physics-driven approach for improving CBCT image quality was investigated to enhance CBCT-based dose calculations. This approach combines a CBCT data correction pipeline with 2D antiscatter grid for scatter suppression, resulting in quantitative accurate CBCT images. Specifically, the utilization of a 2D antiscatter grid, along with measurement-based residual scatter correction, image lag correction, and beam hardening correction, enables a significant improvement in HU accuracy for linac-mounted CBCT images. Throughout the paper, this approach is referred to as quantitative CBCT (qCBCT).

The accuracy of qCBCT-based dose calculations was assessed by simulating targets in the pelvis and abdomen regions. These targets were treated with hypofractionated radiotherapy regimens using intensity-modulated radiation therapy techniques. The study evaluated the impact of patient size, beam energy, and target location on dosimetric accuracy. Furthermore, the dosimetric accuracy was benchmarked against the gold standard pCT images utilized for radiation therapy treatment planning.

2. Methods

2.1. Quantitative CBCT pipeline

Improving the accuracy of CBCT based dose calculations requires improvement of CT number accuracy. The proposed qCBCT pipeline incorporates several methods to improve the CT number accuracy (Fig. 1). Since one of the major causes of CT number accuracy degradation is scattered x-rays, a two-step scatter suppression approach was implemented. First, a 2D antiscatter grid prototype was integrated onto the flat panel detector, which rejects vast majority of scatter. Previous work has shown that the remaining, not rejected, scatter has detrimental effects on the HU accuracy in pelvis and abdomen sized phantoms, and must be accounted for³². To correct for the residual scatter, the Grid-based Scatter Sampling (GSS) method was implemented^33,34. In addition, image lag inherent to the flat panel detector and beam hardening were also corrected to further improve CT number accuracy³⁵.

Fig.1.

qCBCT pipeline diagram.

The 2D grid prototype in this work was fabricated using the powder bed laser melting technique and consists of a 2D array of tungsten septa aligned towards the x-ray focal spot. It has dimensions of 3 cm in width in the axial direction (parallel to the axis of rotation) and 40 cm in width in the transverse direction. The grid’s focusing geometry was designed for Varian TrueBeam’s offset detector CBCT scan geometry. It features a grid pitch of 2 mm, a grid ratio of 12, and a wall thickness of 0.1 mm. The 2D grid prototype effectively rejects over 90% of the scatter fluence, resulting in improved HU accuracy³⁶.

Image lag was corrected using Mail et al.’s method³⁷, by modeling image lag in flood projections. Water equivalent beam hardening was applied to correct both patient and bow tie filter induced beam hardening in CBCT images³⁸. Bone-specific beam hardening was not implemented as it requires differentiation of bony and soft tissue regions in 3D images first. This approach can be adversely affected by motion due to slow gantry rotation.

For completeness, the GSS method is described below, as it is key to achieving high CT number accuracy with the 2D antiscatter grid approach^32–34. About 3–10% of scatter fluence is transmitted through the 2D grid, which degrades CT number accuracy^32,36,39. The GSS method measures and corrects residual scatter transmitted through the 2D grid, by utilizing the 2D grid as a scatter measurement device. 2D grid wall shadows act as periodic micro modulators of image signal across the area of the detector, where the ratio of image signals in grid shadows and adjacent grid holes change as a function of residual scatter intensity.

In a phantom projection with a 2D grid in place and no residual scatter present, primary signals residing in grid holes, P_hole(u,v), and septal shadows, P_wall(u^’,v^’), in a small neighborhood of pixels (e.g., 2×2 mm²), are expressed as,

$$P_{\mathit{\text{hole}}}\left(u,v\right)G{M}_{\mathit{\text{hole}}}\left(u,v\right)=P_{\mathit{\text{wall}}}\left(u^{\prime },v^{\prime }\right)G{M}_{\mathit{\text{wall}}}\left(u^{\prime },v^{\prime }\right)$$ (1)

Gain map, GM, characterizes signal variations introduced by the 2D grid’s footprint, or wall shadows, in flood projections. Primary signal difference between grid holes and adjacent grid wall shadows are compensated by GM. For any arbitrary pixel (i,j) in a projection, GM is

$$GM(i,j)=\frac{C}{F(i,j)}$$ (2)

F is a flood projection acquired without an object but with a 2D grid on the detector, and C is an arbitrary normalization constant. In this work, C was the average of pixel values in the flood projection. In Eq. 1, it was assumed that primary signal variations due to imaged object can be omitted in a small pixel neighborhood.

When residual scatter is present, it appears as a slowly varying and additive signal. Since GM correction is a multiplicative correction and scatter signal is additive, GM correction would not equalize signals residing in grid holes and adjacent grid shadows. After GM correction, signal difference, d, at a wall shadow location with respect to the adjacent grid hole is,

$$d\left(u^{\prime },v^{\prime }\right)=\left[{\left(P\left(u^{\prime },v^{\prime }\right)+S\left(u^{\prime },v^{\prime }\right)\right)}_{\mathit{\text{wall}}}\right]GM{\left(u^{\prime },v^{\prime }\right)}_{\mathit{\text{wall}}}-\left[{\left(P\left(u,v\right)+S\left(u,v\right)\right)}_{\mathit{\text{hole}}}\right]GM{\left(u,v\right)}_{\mathit{\text{hole}}}$$ (3)

Assuming that scatter signal intensity is piecewise uniform in the grid wall shadow and adjacent grid holes, scatter signal, S, at the grid wall location can be calculated³⁴ from Eqs. 1 and 3

$$S\left(u^{\prime },v^{\prime }\right)=\frac{d(u^{\prime },v^{\prime })}{{GM}_{\mathit{\text{wall}}}\left(u^{\prime },v^{\prime }\right)-〈{GM}_{\mathit{\text{hole}}}(u,v)〉}$$ (4)

$〈{GM}_{\mathit{\text{hole}}}〉$ is the average of GM values in grid holes in the small pixel neighborhood surrounding the pixel of interest residing in a wall shadow.

The signal difference, d, was calculated between each pixel residing in wall shadows and neighboring grid holes in a projection in a small neighborhood (2×2 mm²). This process was repeated for all pixels in the wall shadows. After calculation of d, the value of S was calculated for pixels residing in the grid shadows, and scatter for pixels in the grid holes was calculated via 2D interpolation. After obtaining 2D scatter map for each pixel (i,j), and this map was then subtracted from the phantom projection to correct scatter and obtain the primary-only projection.

$$P\left(i,j\right)=T\left(i,j\right)-S\left(i,j\right)$$ (5)

T(i,j) is the raw phantom projection (primary + scatter). Scatter correction was followed by GM correction to suppress the grid shadows.

$$P_{\mathit{\text{GMCorr}}}(i,j)=P(i,j)\times GM(i,j)$$ (6)

This process was repeated for each projection.

2.2. Acquisition of CBCT images

Based on our experience with HU accuracy evaluations, imaged object composition and size are the two profound factors that affect HU accuracy. Hence, two pelvis phantoms, emulating standard and large body habitus were employed to evaluate the CBCT-dose calculations accuracy. The standard phantom had a lateral and anterior-posterior (AP) separation of 30 and 21 cm, respectively. The large phantom was constructed by adding the Superflab soft tissue mimicking layers around the standard phantom, such that the lateral and AP separation was increased to 42 and 34 cm, respectively. Since the axial field of view was 46 cm in CBCT images, large phantom was designed to fit in the field of view and prevent truncation artifacts.

qCBCT scans were acquired by using the clinical pelvis CBCT protocol parameters: CBCT projections were exported and corrected for residual scatter, image lag, and beam hardening. Subsequently, images were reconstructed by using the FDK method and TIGRE toolkit modified for offset detector reconstruction⁴⁰.

Since the 2D grid had 3 cm width in the axial direction, it provided 2 cm wide field of view in the axial direction in CBCT images (Fig. 2). To image a larger volume of a phantom in the axial direction and be able to calculate dose, 8 contiguous qCBCT scans were performed, where patient couch was shifted in the axial direction between scans. During each scan, radiation field of view on the detector plane covered the full active area of the detector to achieve realistic scatter conditions. These 8 scans were stitched together to achieve 13.6 cm long field of view in the axial direction. When larger 2D grids are fabricated in the future, such multiple CBCT scans would not be needed.

Fig.2:

Pictures of (a) 2D Anti scatter grid and (b) CBCT setup.

In addition to qCBCT scans, two other CBCT configurations were also evaluated, one of them was a clinical CBCT protocol and the other one employed a high performance 1D ASG with a grid ratio of 21³⁶. These two configurations served as references to evaluate the effect of raw data fidelity on the dose calculation accuracy.

Clinical CBCT configuration was the standard pelvis protocol in the TrueBeam system, which employs scatter kernel superposition-based scatter correction, beam hardening correction, couch scatter correction, detector glare correction, lag correction, and a conventional radiographic antiscatter grid with a grid ratio of 10^41,42. Clinical CBCT scans were reconstructed using the FDK method. Whereas CBCT scans acquired with the high performance 1D ASG were not processed with any of the raw data correction methods, and they were reconstructed by using the FDK method as in qCBCT images.

Imaging dose for each qCBCT and clinical scan was the same. 2D grid did not introduce any noticeable increase in image noise. Because, the average primary transmission of the prototype 2D grid in qCBCT scans was 85%, whereas the primary transmission of the conventional antiscatter grid in the clinical CBCT system was 70%^36,43. Therefore, the use of 2D grid in qCBCT scans did not require an increase in imaging dose when compared to clinical CBCT scans.

Both qCBCT and clinical CBCT scans were acquired using the same acquisition parameters as in the TrueBeam pelvis CBCT protocol; Each scan was acquired at 125 kVp and 1080 mAs in offset detector geometry, and with the half-fan bow tie filter in place. Computed Tomography Dose Index (CTDI) was 16 mGy for each scan.

Based on our preliminary evaluations, bow tie filter caused highly heterogenous scatter-to-primary ratios in 1D-ASG projections, leading to severe HU nonuniformities. Therefore, 1D ASG CBCT scans were acquired without the bowtie filter. Scan technique was 125 kVp and 450 mAs. mAs was reduced to prevent detector saturation at the phantom-air boundary due to absence of the bow tie filter.

Planning CT (pCT) scans were acquired at 120 kVp by using a 16 slice Philips Brilliance Big Bore multidetector CT scanner (Philips Medical Systems, Netherlands). pCT images served as the gold standard for treatment plan generation and dosimetric evaluations. CTDI was 16 mGy in pCT scans.

In addition to pelvis phantoms, a large HU to mass density phantom (Gammex advanced electron density phantom, Sun Nuclear Corp, FL), was also scanned using each imaging modality evaluated (Fig. 2b).

2.3. Generation of treatment plans and evaluation of dose calculation accuracy

To assure that targets and organs at risk (OARs) were identical in size and location in all image sets, CBCT and pCT images of pelvis phantoms were coregistered rigidly by using the bony anatomy as reference. A total of 5 targets were delineated in the pCT images (Fig. 3). Two of them were in soft tissues, one emulating a lymph node (LN) proximal to the right iliac wing, the other, a central target and 3 of them in bony regions, with diameters ranging from 2 to 8 cm. To simulate organs at risk (OARs), rind structures with a thickness of 1 cm were placed around each target with a gap of 1 cm between the two.

Fig.3.

5 targets are shown in the coregistered (a) small and (b) large pelvis phantoms. Each target’s corresponding Organ at Risk (OAR), a green structure, is also presented in (a). OARs are rinds at 1 cm distance from their respective targets.

Treatment plans were first generated using the pCT images of pelvis phantoms in the Eclipse Treatment Planning System (Varian Medical Systems, Palo Alto, CA) and the ACUROS Version 15.6 dose calculation algorithm. For dose calculations, HU to mass density tables were first generated from HU-to-density phantom images for each CBCT configuration and the pCT. For each CBCT and pCT configuration, two scans of the HU-to-density phantom were acquired, where the placements of material inserts were changed between the two scans (Fig. 4). This approach aimed to reduce the impact of image artifacts caused by material inserts on the measured HU values. Subsequently, HU values of each respective material inserted in two scans were averaged and used in HU-to-density tables (Fig. 5).

Fig.4.

Two HU-to-density phantom images acquired in each respective scan configuration where the location of material inserts was changed between the two scans. This approach was aimed to reduce location specific HU biases introduced by the material inserts. HU window: [-250 250].

Fig.5.

HU-to- relative mass density conversion tables for all imaging configurations.

HU accuracy in each CBCT configuration and pCT were evaluated by using the HU loss metric,

$$\Delta HU=\left|H{U}_{\mathit{\text{small}}}-H{U}_{\mathit{\text{large}}}\right|$$ (7)

where HU_small and HU_large are the average HU values in the small and large pelvis phantoms for a region of interest (ROI). HU loss was evaluated across 20 bone-mimicking and 18 soft-tissue mimicking ROIs.

In the small pelvis phantom pCT images, one treatment plan was generated for each target that mimicked hypofractionated radiation therapy scenarios, delivering 30 Gy in 3 fractions. Volume Modulated Arc Therapy (VMAT) technique was used for all plans. This process was repeated for 6 and 10 MV flattening filter free (FFF) beams, to evaluate the impact of beam energy on dose calculation accuracy. Since dose calculation accuracy may also depend on VMAT arc direction and length, treatment sites around the left and right iliac wings (Fig. 3) were treated with 210-degree VMAT arcs entering from the left and the right side, respectively. Whereas the central soft tissue and the L5 target were treated with 360-degree arcs. All plans were normalized to deliver 100% of prescribed dose to 95% of the target (D₉₅=100%). This process was repeated for the large pelvis phantom. A total of 20 treatment plans were generated for 5 targets in each pelvis phantom, using 2 different beam energies. After optimization of treatment plans, plans were copied onto the CBCT images of the respective phantoms, and the dose was recalculated using the same treatment plan parameters as in pCT-based plans.

After dose calculations, dose volume histograms (DVHs) were generated. Several dosimetric metrics were calculated for each target, including the mean dose, D₉₅ (which represents the dose delivered to 95% of the volume), V₁₀₀ (representing the volume receiving 100% of the dose), and the maximum dose (the dose delivered to 0.1cc of the target volume). The mean dose and maximum dose covering 0.1 cc were also calculated for OARs. In order to compare the dosimetric metrics obtained from different CBCT modes, the pCT-based values were assumed to be 100%, while the metrics obtained from other CBCT modes were normalized to their corresponding pCT-based values. Lastly, the error was defined as the difference between the DVH metrics of the relative CBCT-based treatment plan and the corresponding pCT-based plan.

3. Results

Images of the pelvis phantoms are shown in Fig. 6, which qualitatively demonstrate the HU variations in all data sets investigated for bone-mimicking ROIs. qCBCT provided the lowest HU loss among the 3 CBCT modalities. (Fig. 7). Median HU loss for soft tissues in qCBCT and pCT images were 7.7 and 3.2 HU, respectively, implying that HU accuracy in soft tissue regions were comparable in qCBCT and pCT images. In bony regions, median HU loss for qCBCT and pCT was 31 and 46 HU, respectively, indicating that HU values in qCBCT were more accurate than the ones in pCT in bony regions.

Fig.6.

Images of small and large pelvis phantoms generated by all imaging methods investigated. HU window: [-200 1000] is selected to demonstrate HU variations in bone mimicking ROIs.

Fig.7.

HU loss as a function of imaging methods in bony and soft-tissue ROIs. Central mark indicates the median, and the bottom and top edges of the box indicate the 25th and 75th percentiles, respectively. Whiskers extend to the most extreme data points.

Similar to the trends in the HU loss, qCBCT and pCT based dose distributions were in agreement with each other. An example is shown in the axial and sagittal (Fig. 8) views of the L5 target treated with a 6 MV beam. Such dosimetric agreement between qCBCT and pCT was also evident in the DVH plots of 6 MV and 10 MV plans for both bony (Fig. 9 and Table 1) and soft-tissue (Fig. 10 and Table 2) targets.

Fig.8.

Calculated Isodoses for the L5 vertebral body target in both a transverse and sagittal slice in the large and small pelvis phantoms with 6MV beam energy.

Fig.9.

DVH plots for the L5 Target treated with the 6 MV beam (a) for the small and (b) large pelvis phantoms. (c) and (d) 10 MV beam with a small pelvis phantom and 10 MV beam with a large pelvis phantom.

Table 1.

Mean and standard deviation of DVH metrics difference with respect to pCT (%) for L5 target in treatment plans presented in Fig. 9.

	D95 (%)	Mean dose (%)	Dmax (%)	V100 (%)
1D ASG CBCT	2.6±2.1	2.6±2.4	3.5±2.6	4.2±3.3
Clinical CBCT	2.4±3.1	1.3±0.9	1±0.5	1.6±1.7
qCBCT	0.1±0.1	0.2±0.1	0.3±0.09	0.1±0.09

Fig.10.

DVH Figures for Lymph Node Target using (a) 6 MV beam with a small pelvis phantom (b) 6 MV beam with a large pelvis phantom (c) 10 MV beam with a small pelvis phantom and (d) 10 MV beam with a large pelvis phantom.

Table 2.

Mean and standard deviation of DVH metrics difference with respect to pCT (%) for the lymph node target in treatment plans presented in Fig. 10.

	D95 (%)	Mean dose (%)	Dmax (%)	V100 (%)
1D ASG CBCT	3.9±1.9	4.4±1.1	4.5±1	21.1±21.1
Clinical CBCT	1.4±1.4	0.8±0.6	0.6±0.4	1.7±1.4
qCBCT	0.15±0.03	0.1±0.05	0.1±0.07	0.2±0.04

While dosimetric errors were relatively small in the 1D ASG and Clinical CBCT based dose-calculations for the small pelvis phantom, dosimetric discrepancies increased in the large pelvis phantom. When compared to pCT based plans, dose was underestimated in 1D-ASG CBCT and overestimated in Clinical CBCT images.

Overall, phantom size, beam energy, target location, and tissue type had small impact on the dosimetric errors in qCBCT dose distributions. Among the DVH metrics investigated across all plans, the median and maximum errors in D₉₅ were 0.1% and 0.2% in qCBCT treatment plans, respectively (Fig. 11). Dosimetric errors in V₁₀₀, D_max, and mean dose were similar across all qCBCT based dose distributions. In the small pelvis phantom, maximum DVH metric error was 0.4% across all DVH metrics investigated.

Fig.11.

DVH metrics difference with respect to pCT for all targets and setups.

On the other hand, median and maximum dosimetric errors in the clinical CBCT based plans were 0.57% and 7% across all DVH metrics, indicating stronger dependence of dose calculation accuracy on beam energy, tissue type, and phantom size. Phantom size was the leading factor in dosimetric errors. The largest dosimetric errors were observed in the large pelvis phantom, in bony targets, and when using 6 MV beams. In the large phantom, reducing the beam energy from 10 to 6 MV increased the maximum error in D₉₅ from 1.1% to 7% for bony targets (Fig. 12). Whereas maximum error in D₉₅ for soft tissue targets was 2.3% when using 6 MV beams. Dosimetric error differences between central and lateral targets (such as L5 target and left iliac wing target) were small, implying that geometric location of the target and beam orientation have a small effect on the dosimetric accuracy.

Fig.12.

DVH metrics difference with respect to pCT for targets from (a) 6 MV and (b) 10 MV beam energy plans.

1D ASG CBCT based plans exhibited even larger dosimetric errors. Median and maximum errors among all target DVH metrics and plans reached 4% and 97% respectively. Maximum error in V100 ranges from 54% to 97% based upon phantom size (Fig. 13). Changing the energy from 10 to 6 MV increased the errors in V₁₀₀ from 5% to 36% for bony targets (Fig. 14). Trends in D₉₅, D_max, and mean dose for targets were similar to the ones observed in V₁₀₀.

Fig. 13.

DVH metrics difference with respect to pCT for targets in the (a) small pelvis and (b) large pelvis phantoms for all beam energies and target tissue types combined.

Fig. 14.

DVH metrics difference between CBCT and pCT based dose calculations for (a) soft tissue targets and (b) bony targets for all phantom sizes and beam energies combined.

Median and maximum DVH metric errors for OARs among all qCBCT-based plans, phantoms, and possible metrics were 0.11% and 0.31%, respectively (Fig. 15). While OAR dosimetric errors were higher in clinical CBCT plans, maximum error in D_max was less than 3% among all plans. 1D ASG CBCT plans exhibited the largest dosimetric errors in OARs, where maximum error in D_max was 7%. When compared to targets, the effects beam energy, tissue type, and phantom size had less noticeable impact on the OAR dosimetric errors.

Fig. 15.

CBCT DVH differences with respect to pCT for OARs of (a) soft-tissue and bony targets, (b) targets treated with 6 and 10 MV beams and (c) targets in small and large pelvis phantoms

4. Discussion

There has been much discussion about improving the accuracy of CBCT-based dose calculations⁴⁴. Several different approaches have been proposed. (1) HU to density table generation can be tailored to patient size^12,45. This method reduces the patient size dependent HU inaccuracy, but it is not sufficient to achieve high dosimetric accuracy^10,12. Moreover, having multiple HU to density tables requires a decision step to pick the right table, which increases the complexity of the workflow. Selection of suboptimal HU to density table may even exacerbate the dosimetric errors. (2) Organ or anatomy specific HU override in CBCT images provides improvement in dose calculation accuracy, but this approach adds one more step to online dose calculation process^10,46,47. Deformable image registration (DIR) may help to accelerate this process, where HU values in pCT images are mapped onto CBCT images. However, errors in deformable registration can propagate to errors in online dose calculations^22,48. (3) A variant of the latter approach is to map dose values in pCT images to CBCT images, through deformable image registration.^22,49 However, this approach would not take major anatomical changes into account during dose estimations, such as weight loss, and significant reduction in tumor volume. Generation of CT-like images from CBCT using machine learning methods is another approach to improve accuracy of HU values and online dose calculations^23–27,50. While these methods provide promising results, the differences between the training datasets and actual radiotherapy imaging conditions may degrade their performance.

Whereas our approach utilized physics-driven methods to improve the CT number accuracy in CBCT images, which in return, increased the dose calculation accuracy in CBCT images. This work demonstrates the importance of robust scatter suppression in CBCT for dose calculations in pelvis, enabled by the 2D antiscatter grid and the GSS scatter correction method. In essence, the dosimetric differences between the planning CT and qCBCT images were not significant, implying that qCBCT can potentially provide planning CT-like dose calculation accuracy using CBCT images acquired before or during radiation treatment delivery.

It is important to emphasize that small dosimetric discrepancies between pCT and qCBCT based plans cannot be strictly classified as dosimetric errors in qCBCT based plans. As shown in the HU loss plot in Fig. 7, the CT number accuracy of pCT was lower than qCBCT in bony regions, which in turn may reduce the dosimetric accuracy for bony targets in pCT-based dose calculations. Even though experiments and analyses were focused on CBCT-based dose calculations for targets in the pelvis region, similar dosimetric accuracy is expected in the abdomen region due to comparable dimensions and tissue composition in the abdomen and pelvis regions.

For reference purposes, the dose was also calculated in CBCT images utilizing a high performance 1D ASG. While high performance ASG was not sufficient by itself to achieve highly accurate dose calculations, results obtained with it demonstrate the severity of potential dosimetric errors that may occur without robust raw data correction strategies.

Clinical CBCT images provided substantially higher dosimetric accuracy due to improved scatter suppression when compared to the 1D ASG CBCT. Dosimetric differences between the gold standard pCT and Clinical CBCT dose calculations were small in most instances, particularly in the small pelvis phantom, regardless of the beam energy and treatment site. This is because the size of the pelvis phantom was similar to the CT to density phantom dimensions, and thus, HU values for a given tissue type were expected to be comparable in both phantoms. However, CT number accuracy degraded substantially in the large phantom as demonstrated in the HU loss plots. As a result, dosimetric errors were substantially larger for targets in the large pelvis phantom.

Dosimetric errors in clinical CBCT and 1D ASG CBCT based plans can be potentially reduced by matching the size of the HU-to-density phantom to pelvis phantom size. However, this approach requires storing multiple HU-to-density tables in the treatment planning system, and manually selecting the HU-to-density table during treatment planning phase, based on the imaged object size. If HU-to-density table is not correctly matched to the object size, it may increase dosimetric errors. Therefore, this approach was not investigated in our study.

Several areas regarding the dosimetric accuracy of qCBCT based plans remain to be investigated in future studies. First, our study compared the dose calculation accuracy with respect to clinical CBCT images that employed scatter kernel superposition-based scatter correction⁵¹. More recently developed clinical CBCT methods utilize more accurate model-based scatter correction algorithms and achieve more accurate CT numbers⁵². This more advanced clinical CBCT imaging method was not available to the authors at the time of this study. Second, this work was focused on the dose calculation errors in the pelvis-abdomen region. Dose calculations in thorax, where HU values are highly heterogenous, may benefit from more accurate qCBCT-based dose calculations. Besides photon dose calculations, CBCT-based proton dose calculation is an area of interest, where errors in CBCT HU values can result in proton range uncertainties. Therefore, improved HU accuracy by qCBCT can potentially translate to larger gains in dosimetric accuracy in proton therapy.

5. Conclusions

CBCT-based dose calculations can play an important role in verification of treatment dose delivery and online modification of treatment plans to assure intended dosimetric coverage of targets and sparing normal tissues. Combination of 2D antiscatter grid with raw data correction methods in the qCBCT approach can provide highly accurate CT numbers thereby allowing accurate radiation treatment dose calculations in the pelvis and abdomen region. This approach may negate the need for density overrides or co-registration of planning CT and CBCT images, thereby improving the clinical workflow and making CBCT-based dose calculations clinically practical.