Per-Fraction Positional and Dosimetric Performance of Prone Breast Tangential Radiotherapy Assessed with Daily Rapid Kilo-voltage Cone Beam Computed Tomography: Key Statistics, Predictors, and Recommendations

This study investigates daily breast geometry and delivered dose to prone-positioned patients undergoing tangential whole breast radiation therapy (WBRT) on an O-ring linear accelerator with 6X attening lter free mode (6X-FFF), planned with electronic compensation (ECOMP) method. Most practices rely on skin marks or daily planar image matching for prone breast WBRT. This system provides low dose daily CBCT, which was used to study daily robustness of delivered dose parameters for prone-positioned WBRT.


Introduction
Breast cancer is diagnosed in about a quarter of a million women in the United States every year [1]. Lumpectomy with tangential-eld, whole-breast radiotherapy (WBRT) is a standard treatment as part of breast conservation therapy (BCT) for early stage disease [2,3]. Tangent WBRT can be delivered in either supine or prone position. Prone position may be optimal for patients with larger and more pendulous breasts for three reasons: (1) Reduced dose to heart and lungs [4], (2) respiratory motion of the chest walls and surgically inserted clip is reduced, consequently decreasing intra-fraction dosimetric variations [5], and (3) dose inhomogeneities in supine position raises the risk of late skin effects resulting in adverse cosmesis.
Daily image guidance is crucial in prone patient positioning, since immobilization devices alone cannot provide consistent day-to-day breast localization. Matching clips or bony / soft tissue anatomy with orthogonal electronic portal image devices (EPID) is a common standard for image guidance in prone WBRT. EPID positional accuracy is comparable to kV cone-beam CT (CBCT) guidance [6,7]. Unlike EPID, however, CBCT can provide tomographic image from which 3-dimensional positional and dosimetric outcomes can be derived. Despite this advantage, busy clinics are discouraged from adopting daily CBCT because CBCT acquisition time may decrease throughput and increase imaging dose.
Halcyon™ v2.0 (Varian Medical Systems, Palo Alto, CA) is a commercially available 6MV attening-lterfree O-ring linear accelerator that provides rapid daily kV CBCT image guidance for tangential WBRT. The O-ring design reduces the risk of collisions and enables higher gantry rotation speed for faster CBCT imaging, which demonstratively improved patient throughput in many clinical sites including whole breast [8,9]. Compared to C-arm linacs, Halcyon v2.0 CBCT acquisition is faster (17-42 seconds versus 60 seconds), even after taking into account reconstruction duration for iterative CBCT, or iCBCT [10], which results in a better contrast-to-noise ratio.
One trade-off for higher patient throughput with Halcyon is the restriction to 6X-FFF beams, which results in a non-at dose pro le at depth. Electronic tissue compensation (ECOMP) planning technique [11] is a forward-planning technique that compensates for this non-at pro le to homogenously treat an irregular surface (i.e. breast) using parallel opposed beams. In contrast, Field-in-eld (FiF) is a more common planning technique for C-arm linacs. It has been previously demonstrated that FiF takes signi cantly longer on Halcyon (9 min FiF vs. 3-4 min ECOMP) owing to Halcyon's dual multi-leaf collimator (MLC) system.
It is unknown if ECOMP with FFF beam for prone WBRT meets dose coverage metrics (i.e. V90%) or hotspot metrics (i.e. V105%) on a per-fraction basis. American Society for Radiation Oncology (ASTRO) guidelines on whole breast RT state that V105% (volume of breast receiving 105% of prescription dose) should be minimized at all times [12], as V105% has been linked to adverse cosmesis [13]. ASTRO guidelines also state that V95% should cover the whole breast fully. The rapid CBCT capabilities of Halcyon v2.0 enables large-scale studies of dosimetric robustness, for example, by deforming planning CT to daily CBCT. Such capabilities can guide adaptive radiotherapy decisions if needed should plans fail to meet the guidelines.
In this paper, we assess the daily positional and dosimetric quality of ECOMP with FFF beam for tangential-beam prone WBRT, using per-fraction CBCTs provided by Halcyon v2.0 during patient positioning to: (1) Assess daily dosimetric and positional robustness of our treatments using daily CBCTs in real patient cases, and (2) investigate what factors affect dose homogeneity most adversely.
Assessment of these factors may possibly inform future adaptive RT decisions.

Methods
Overall Design of the Study The purpose of this study is to evaluate daily positional and dosimetric accuracy of tangential-eld prone WBRT, planned with ECOMP and treated with 6X-FFF beam on Halcyon™ version 2.0 (Varian Medical Systems, Palo Alto, CA). We seek to investigate what parameters affect target coverage and radiation hotspot the most, using Halcyon's rapid daily CBCT capabilities. This retrospective study was reviewed and approved by our Institutional Review Board.
Image acquisition was done as follows. Eight (8) prone breast patients receiving WBRT post-lumpectomy with hypofractionated schedule of 2.66 Gy x 16 fractions followed by 10 Gy boost were retroactively analyzed. CT simulation was performed after patients were immobilized on a commercially available prone breast board. A kV CBCT was taken before delivery of each fraction, as part of Halcyon v2.0 imageguided RT work ow. A total of 16 fx * 8 patients = 128 CBCT datasets were obtained. A synthetic daily CT was generated by deformably registering the planning simulation CT to each of the daily CBCT images using MIM Maestro® (MIM Software Inc., Cleveland, OH) version 6.6. CBCTs associated with boosts were excluded from daily evaluations.
The clinical treatment plans were generated using ECOMP technique (EZ uence™, Radformation, Inc, New York, NY) in Eclipse v15.6 (Varian Medical Systems, Palo Alto, CA), with dosimetric endpoints speci ed in Table 1. This plan was copied and recalculated on synthetic CT to obtain daily delivered dose distribution. BODY and Breast PTV_eval contours, as de ned on Radiation Therapy Oncology Group (RTOG) trial 1005 [14] and RTOG census de nitions [15] were created on each synthetic CT. The Breast PTV_eval contour was derived for each day of treatment, the process of which is clari ed in the following subsections.

Dosimetric Endpoints Investigated
Most recent ASTRO guidelines on WBRT state that V105% should be minimized at all times, while some cite 200 cm 3 as the recommended limit [12]. The tumor bed is to receive at least 95% of the prescription dose under ASTRO guidance. Similarly, RTOG1005 protocol stipulates (1) 95% of Breast PTV Eval shall receive at least 95% of prescribed dose (D95% > 95%, but D90% > 90% acceptable), and (2) maximum dose less than 115% of prescribed dose (Dmax < 115%, but 120% acceptable).
To investigate if delivered prone WBRT dose on Halcyon meets these guidelines on a per-fraction basis, three primary dose-volume metrics were measured at each fraction to assess PTV Eval coverage and hotspot: V90%, D95%, and V105%. Absolute values of V95% [cc], V100% [cc], and V105% [cc], and global maximum dose (Dmax) were obtained from the BODY contour. Predictive model for these dose-volume metrics were constructed, based on patient parameters including source-to-surface distance (SSD), couch shifts, residual breast position (de ned below in Data Collection, Analysis, and Statistics section), and weight.
Patient Set-up and Initial Treatment Planning Eight (8) patients were simulated head-rst-prone on Siemens Sensation CT scanner (Siemens Healthineers, Erlangen, Germany) for initial treatment planning. Patients were immobilized with QFix® Prone Breast (Avondale, PA) boards and a vac-lok bag placed underneath the patient and with their arms up. Imaging isocenter was placed midline of the body, at the midpoint (sup-inf) of the breast tissue and anteriorly to the sternum.
WBRT was planned in Varian Eclipse v15.6. Physicians contoured Breast CTV based on consensus de nitions [15]. Breast CTV was expanded 5 mm (excluding heart and not crossing the midline) and cropped anteriorly from the skin by 5 mm and posteriorly in front of the rib to obtain Breast PTV Eval.
Treatment isocenter was shifted anteriorly to cover Breast PTV Eval. ECOMP technique was carried out by experienced dosimetrists. ECOMP is a forward-planning intensity-modulated radiation therapy (IMRT) technique using parallel opposed beams, where the goal is to deliver as homogenous dose as possible to an irregular surface. For breast, this is done by planning a uniform dose at mid-separation. Skin ash of 2 cm was added beyond the patient contour, and then the edited uence maps were converted to leaf sequences for Halcyon dynamic MLC. Machine energy was xed at 6X-FFF (6-MV with attening lter free), with prescription dose of 266 cGy/fraction for all patients. Anisotropic analytical algorithm (AAA) version 15.6.03 was used for volumetric dose calculation. For photon dose optimization and irregular surface compensator generation, photon optimizer version 15.6.03 was used.
Dose-volume constraints for Breast PTV Eval in a whole breast treatment used at our institution are listed on Table 1. D95% and V90% measure the extent of dose coverage to the breast, while V105% and Dmax are measures of dose hotspots. For this study, a < 200 cm 3 objective is additionally applied for the entire patient volume (BODY contour) as another suggested measure of homogeneity by ASTRO consensus but was not a constraint used clinically during planning. Study of normal organs, such as the lungs, hearts, and contralateral breast were omitted from this study. CT Deformation to Daily CBCT and Daily Treatment Planning Figure 1A demonstrates the process to generate daily synthetic CTs using MIM Maestro v6.6. In the MIM console, planning CT was deformably registered to the CBCTs to generate synthetic daily CTs. Some synthetic CTs were discarded (2 in total) because of deformation failures stemming from limited CBCT eld-of-view. A total of 126 synthetic CT images was extracted from MIM and imported back into Eclipse. Once imported, a copy of original RT plan was re-calculated on each of the synthetic daily CTs keeping all beam parameters intact, while taking into consideration the online CBCT-to-CT registration that was performed during daily treatment. Dose to synthetic daily CTs was calculated with the same AAA version as the original treatment plan. Figure 1B demonstrates the contouring process on daily synthetic CTs for the treatment target, which is Breast PTV Eval contour. Original Breast PTV Eval contour (from the planning CT) were rst copied into each daily synthetic CTs keeping the posterior edge shape and position relative to the chest wall the same as in the initial plan. The transferred PTV Eval was then modi ed using the same guidelines that was used to generate PTV Eval during the initial planning: a 5 mm subtraction from the breast tissue surface. The breast tissue surface was automatically determined by the Eclipse software in the contouring module for each of the synthetic CTs. Overall, this procedure ensures the PTV Eval structure generated on daily synthetic CT consistently represents the initial physician's intent on breast tissue being irradiated for each daily fraction.

Data Collection, Analysis, and Statistics
We make a distinction between two categories of data: (1) planning data and (2) per-fraction / daily data. The former refers to the baseline parameters determined from planning CT and the planned dose. The latter refers to the treatment parameters determined from daily synthetic CTs and the delivered dose. The difference of daily data from planning data is denoted throughout this paper with an uppercase Greek delta (Δ). For example, ΔV105% = Daily V105% -Planned V105%. Other data collected in this study include machine data: collimator angles, gantry angles, and monitor units for each of the two parallel opposed beams. Patient weight data gathered during simulation and physician visits was spline-interpolated to estimate weight at every fraction.
All data collected were imported into MATLAB (Natick, MA) Version 2019a for data visualization and statistical analyses. Tables and histograms of predictor variables, dosimetric endpoints, and patient position information are presented with descriptive statistics. Non-parametric statistical tests (Wilcoxon signed-rank test) were used to compare daily delivered dose-volumes with dose-volume optimization criteria (Table 1) and with planned dose-volumes. Pearson correlation coe cients among all variables studied are presented, with emphasis on statistical signi cance of the observed correlation.
Finally, robust parallel slopes model (a type of analysis of covariance or ANCOVA) was constructed [16] to t Δ dosimetric endpoints using the following predictors: couch shifts, PTV breast residual shifts, Δ BWD, 1-DSC, Δ PTV Eval Volume, and Δ body weight. In this model, each patient is assigned an intercept but the slopes with respect to predictors are assumed to be equal, hence the term "parallel slopes." The general tted formula was: Here, is the Δ dosimetric endpoints (e.g. ΔV105%), is the intercept for i th patient, is the slope for n th predictor , and is a normally distributed residual term. The individualized intercept for each patient attempts to compensate for patient-speci c factors, such as the differences in optimized IMRT elds, machine parameters (gantry angles, monitor units), and other miscellaneous factors. A reasonable estimate of the slopes was made based on this model. Robustness of the model was improved with MATLAB's default iterative robust least squares procedure using bi-square weights. This reduces effects of high-leverage data and outliers on the t. Wald's Test with the null hypothesis that estimated statistical signi cance of each variable on the Δ dosimetric endpoints.
Signi cance was set at α = 0.05 for most statistical analyses except for multiple tests, for which Benjamini-Hochberg procedure was applied to limit false discovery rate [17].  Abbreviations: cc = centimeter cubed (cm 3 ); PTV Eval = planning treatment volume for evaluation; LAT = lateral; VRT = vertical; LNG = longitudinal; MAG = magnitude; SSD = source-to-surface distance; DSC = dice similarity coe cient; Stdev = standard deviation.  [cc] was found to be < 200 cm 3 . Median values of all quantities were within the variation acceptable (see Table 1), but some were not within the treatment objective. Figure 2 shows histograms of daily delivered dosimetric outcomes and difference from prescribed (Δ). The number of treatments that meet the prescribed dosimetric endpoints (i.e. "pass") or has a small Δ (e.g. "<10%") are represented as a fraction out of all treatments (126) examined. The histograms corroborate that target coverage metrics were stable, but dose hotspot metrics were not. Dose objective D95% > 95% was met 111 days out of 126 daily CTs examined, while V90% > 99% was less frequently met at 74/126 (but 95/126 within variation acceptable, V90%>95%). For difference from prescribed, ΔD95% was >-5% for 123/126 days and ΔV90% was >-1% for 111/126 days. BODY V90% deviated only slightly from planned, with a standard deviation of ± 51.0 cc with a bias of -16.3 cc. In contrast, PTV V105% < 10% was met only 50/126 days (68/126 within variation acceptable). Dmax was the most often not met, with only 21/126 days meeting the criteria of < 107% of the prescribed dose.  Patient body characteristics such as weight, SSD-derived BWD, and breast volume may explain elevated V105% and Dmax in some of these patients. For example, on Fig. 5, intra-patient ΔDmax increased by 0.53% and PTV ΔV105% increased by 4.6% for every 1% weight loss from CT simulation (reminder that Δ signi es difference from prescribed). Patient 2 lost ~ 4% of her weight since CT simulation even before treatment began and 2% more during the course of the treatment. It is possible that dose delivered to patient 2 failed to meet dose constraints due to the signi cant weight loss.

Descriptive Statistics of Positional and Dosimetric Data
Weight loss likely correlates with ΔV105% and ΔDmax because it decreases breast deformation (DSC) and BWD. In our comprehensive parallel slopes model (Additional le 1), reduced DSC and BWD resulted in increased ΔV105% and ΔDmax. Since weight loss decreases BWD and DSC ( Fig. 4 and Fig. 5B), weight loss should increase ΔV105% and ΔDmax according to the parallel slopes model. Interestingly, weight was not a signi cant predictor of ΔV105% and ΔDmax when both BWD and DSC was present. It is possible that including these two variables in the model effectively cancel out the in uence of weight loss on ΔV105% and ΔDmax. DSC and BWD reduction from weight loss were not the sole cause of ΔV105% and ΔDmax increase. In our model, PTV eval volume also signi cantly in uenced intra-patient variability in ΔV105% and ΔDmax.
Surprisingly, weight did not correlate with PTV eval volume (Fig. 4). This is interesting since breast volume is expected to decrease with weight loss. It is possible that post-lumpectomy seroma resolution is decoupling breast volume from weight loss. Indeed, PTV eval volume as a whole was also observed to decrease with time (r = -0.2) in Fig. 4, suggesting gradual seroma resolution over the treatment course.
In our model for dose coverage metric PTV ΔV90%, BWD, weight, and PTV Eval volume in uences dose coverage along with some positional shifts (namely: vertical couch shift, longitudinal residual shift, and magnitude of residual shift). It is unclear why these speci c couch and residual shifts decrease PTV ΔV90%. It is possible that negative vertical couch movements (i.e. couch moves down) increase PTV ΔV90% since that encourages more of the PTV Eval Breast to be within the eld. It is also possible that residual shifts decrease PTV ΔV90%, because any post-couch-shift residual movement of the breast will knock the breast out of the con nes of the parallel-opposed eld. Nevertheless, the slopes of these predictors with respect to PTV ΔV90% are quite small compared to PTV ΔV105%, again corroborating that ΔV90% is robust.
A general guideline to reduce hotspots can be established based on the conclusions from parallel slopes model to ensure dose hotspot is limited and dose coverage is maximized. First, patient's weight and breast volume should be monitored closely for any extreme changes. Second, if patient weight loss or breast volume changes is over a certain threshold, BWD and breast volume should be re-examined during daily CBCT. A re-plan should be considered if BWD or breast volume changed signi cantly and would result in excessive hotspots or a signi cant loss in dose coverage.
The following are preliminary recommendations based on the estimated parameters from the model established on Additional les 1 and 2 and on the dose-volume objectives laid out on Table 1 during the course of the treatment: During CT simulation and planning, measure breast width diameter and extract breast volume from Breast PTV (not Breast PTV Eval).
If a patient lost weight by > 3% compared to during simulation, re-measure breast width diameter and breast PTV volume of the patient in prone position henceforth, even if the patient's weight normalizes.
If a two-day average of either patient's breast width diameter or breast volume deviates more than 1.5 cm or 200 cc compared to during CT simulation, respectively, consider a re-plan for the subsequent fractions.
The goal of the recommendations above is to limit per-fraction PTV ΔV105% (change from planned) to < + 15%, without signi cantly affecting clinical work ow and treatment time. Note that, because BWD is derived from SSD, it is subject to set-up errors. The two-day average helps smooth out the daily variations due to set-up errors. Weight is a metric that is easy to measure during weekly on-treatment visits and can be obtained more frequently easily. Since 1% weight loss results in 4.6% increase in PTV ΔV105%, 3.3% weight loss corresponds to PTV ΔV105% of + 15% (rounded down to 3% Note that, if weight cannot be measured regularly, one can skip measuring the weight and monitor the two-day averages of breast width diameter (as de ned in Fig. 2) and volume daily.
There are several limitations with the model despite the high adjusted R 2 reported (0.94 for V105% and 0.84 for V90%). The intercept estimates (Additional le 2) show inter-patient biases are quite signi cant for V105% and Dmax that cannot be fully explained by the slope components of the regression. It is expected that all slopes should be zero if there were no patient-to-patient biases that predisposes one patient to receive higher or lower dose. There may be second order effects, interaction effects (i.e. variable slopes per patient rather than a single slope), or simply effects from personalized optimized IMRT elds that explain the non-zero intercepts. Another explanation could be that dosevolume metrics like V90% and V105% are inherently non-linear. Whatever the case may be, slopes on Additional le 1 should only be used to estimate intra-patient variations (i.e. fraction to fraction), never inter-patient variations.
Another limitation is that the model is statistically low-powered (n = 126, but only 8 patients) and fails to explain a few outlier or non-linear behaviors. In addition, despite the use of robust methods, our results may have been affected by patient 2, who experienced disproportionately severe weight loss than other patients in the trial. There were outlier behaviors the model did not predict as well, such as signi cant under-dosing (i.e. low PTV V90%) close to the chest wall for some fractions. For the two cases where BODY V105% [cc] was exceptionally high, there was an unexpected high-dose area near the arm outside of PTV for one patient (P4, fraction 7) and a signi cantly decreased breast size for another (P5, fraction 15).

Conclusion
We examined and modeled the daily behavior of dose homogeneity for tangential eld prone-positioned WBRT, using daily Halcyon v2.0 kV CBCT. The goal was to assess day-to-day robustness of electronic tissue compensation (ECOMP) technique with 6X-FFF, a modality that is unique to Halcyon. While metrics of dose coverage such as V90% was often met, metrics of dose hotspot such as V105% and Dmax routinely exceeded speci ed dose constraints on a per-fraction basis for select patients. Data suggests inter-patient differences in dose hotspot could come from patient body characteristics, such as weight, breast volume, and SSD-derived breast width diameter (BWD). In the comprehensive model, three factors explained most intra-patient variations in V105% and Dmax: breast volume, BWD, and breast deformation compared to planned. For some patients, severe weight loss affected breast width diameter and breast deformation, resulting in higher V105%'s than planned. Some patients also experienced breast volume β 0 i β n losses independently of weight, possibly due to seroma resolution, that signi cantly increased V105% and moderately increased Dmax. Data suggests that some patients vulnerable to extreme weight loss or breast volume changes from seroma resolution may bene t from re-planning.

Funding
The authors state that this work has not received any funding. shift from matching CBCT-CT was applied to copied plan isocenter to correct for daily couch shifts. B.
Generated Daily Breast PTV Eval mimicked the original contour as closely as possible, including a 5mm subtraction from the skin surface. C. Some predictive variables used in this paper are de ned here. SSDderived breast width diameter (BWD) is calculated from source-to-surface distances (SSD) of lateral and medial beams, giving an estimate of patient's breast diameter as well as any set-up errors resulting in distorted SSDs. Residual shift vector was derived from comparing centers of mass of planned breast PTV eval volume to that of daily volume, adjusted for couch shift. Dice similarity coe cient (DSC) was calculated from this same setup using the intersection and union of PTV volumes.
the Breast PTV Eval contour. Bottom row: for BODY contour. Number of treatments that meet the dosevolume objectives laid out on Table 1 is represented as a fraction out of 126, in the following format 'N/126 (M/126)' where N = # meeting objectives and M = # within the variation acceptable. B.
Corresponding histograms of daily dosimetric outcome, represented as difference from what was originally planned for the treatment courses. Number of treatments that do not deviate signi cantly from planned (N) are represented in format 'N/126', where "signi cant deviation" means "minimum deviation that would cause a plan meeting all dosimetric endpoints to extend beyond variation acceptable.   Body weight correlates with DSC, BWD, and PTV dose homogeneity and hotspot metrics. Data from patients who consistently lost weight over the course of the treatment are denoted in red (P2, P4, P5, and P6), while others are denoted blue. Linear ts (dotted lines) with equations and regression coe cient are also shown. A. Some patients were susceptible to weight loss, with one patient (P2) losing 4% body weight from simulation even before treatment began. B. DSC and BWD correlated to body weight. C.
Breast PTV Eval ΔV90%, ΔDmax, and ΔV105% (Δ = change from planned) correlates well with Δ body weight. Weight loss correlated to decreased PTV V90%, increased Dmax, and increased V105%, highlighting its crucial role in target coverage and reduction of hotspot.