Contour‐based lung dose prediction for breast proton therapy

Abstract Purpose This study evaluates the feasibility of lung dose prediction based on target contour and patient anatomy for breast patients treated with proton therapy. Methods Fifty‐two randomly selected patients were included in the cohort, who were treated to 50.4–66.4 Gy(RBE) to the left (36), right (15), or bilateral (1) breast with uniform scanning (32) or pencil beam scanning (20). Anterior‐oblique beams were used for each patient. The prescription doses were all scaled to 50.4 Gy(RBE) for the current analysis. Isotropic expansions of the planning target volume of various margins m were retrospectively generated and compared with isodose volumes in the ipsilateral lung. The fractional volume V of each expansion contour within the ipsilateral lung was compared with dose–volume data of clinical plans to establish the relationship between the margin m and dose D for the ipsilateral lung such that V D = V(m). This relationship enables prediction of dose–volume VD from V(m), which could be derived from contours before any plan is generated, providing a goal of plan quality. Lung V 20 Gy( RBE ) and V 5 Gy( RBE ) were considered for this pilot study, while the results could be generalized to other dose levels and/or other organs. Results The actual V 20 Gy( RBE ) ranged from 6% to 23%. No statistically significant difference in V 20 Gy( RBE ) was found between breast irradiation and chest wall irradiation (P = 0.8) or between left‐side and right‐side treatment (P = 0.9). It was found that V(1.1 cm) predicted V 20 Gy( RBE ) to within 5% root‐mean‐square deviation (RMSD) and V(2.2 cm) predicted V 5 Gy( RBE ) to within 6% RMSD. Conclusion A contour‐based model was established to predict dose to ipsilateral lung in breast treatment. Clinically relevant accuracy was demonstrated. This model facilitates dose prediction before treatment planning. It could serve as a guide toward realistic clinical goals in the planning stage.


| INTRODUCTION
Radiation dose distribution depends on target size and target location relative to organs at risk. Therefore, it is often difficult to know a priori what the dose to normal tissues for a patient may be achievable. This can make selecting the optimal treatment approach for a specific patient challenging. Being able to predict the dose to normal tissues prior to treatment planning can guide radiation oncologists and physicists toward realistic clinical goals in the planning stage and ensure the quality of treatment plans. This could also enable physicians at hospitals without proton capabilities to make a betterinformed referral decision or aid patient selection.
Modern radiation treatments deliver highly conformal dose to the target. As a result, the falloff isodose lines typically resemble uniform expansions of the planning target volume. We propose to predict isodose lines based on target contour and thus predict dosevolume V D , thereby allowing V D and metrics of plan quality to be derived from contours before any plan is generated.
The treatment of breast cancer is a relatively new application of proton therapy (PT). 1 While recent comparative treatment planning studies of PT for breast cancer have highlighted the significant advantage in heart and lung sparing as well as in target coverage over traditional photon-based radiation techniques, 2,3 for each individual case, it is not straightforward to determine the dose-volume metrics achievable with PT without generating a treatment plan. In this pilot study, we attempt to predict ipsilateral lung (IL) V 20 Gy(RBE) and V 5 Gy(RBE) based on the contours of each individual patient.

2.A | Patient selection
Fifty-two randomly selected patients recently (2016-2017) treated for breast cancer in our clinic with uniform scanning (US) or pencil beam scanning (PBS) proton therapy were included in this study, which has been reviewed by our institutional review board. The diseases were left-sided for 36 patients, right-sided for 15, and bilateral for 1. Eighty percent of the patients had nodal involvement; and partial breast irradiation was not considered in this study. The prescription doses varied from 50.4 Gy(RBE) to 66.4 Gy(RBE).

2.B | Setup and delivery
Computed tomography (CT) scanning was performed for each patient at the simulation with the arms abducted above the head using a custom mold (Alpha Cradle, Smithers Medical Products, Inc., North Canton, OH, USA). Patients were immobilized in the supine position. There was no actual immobilization of the breast. The CTs were performed on a GE LightSpeed/Optima CT scanner.
The target delineation was described in our previous publication. 4 Although controversial in proton therapy, 5 the planning target volume (PTV) is still widely used by clinicians in evaluation of proton treatment plans (e.g., Ref. [4,6,7]). There is also debate over the appropriate use of PTV margins for setup uncertainty and motion for breast cancer. We used a 7 mm margin medially and laterally but not posteriorly. In the distal/posterior direction, the plans were generated to cover the clinical target volumes and evaluated for range uncertainty. 8 In addition, we did not add the margin medially in the supraclavicular region to avoid extending the PTV into the esophagus. For complete review of the uncertainty margins and their application in US and PBS, refer to our recent book chapter. 9 Proton therapy was typically delivered with two anterior oblique en face fields. For US, due to the limitation in field size, matching fields were typically needed to cover the chest wall and supraclavicular nodes, with the match lines feathered. 4

2.C | Dose prediction
We rescaled all the initial prescription doses to 50. Ten patients were randomly selected, for whom m was adjusted to minimize the root-mean-square error (RMSE) of prediction V(m) where the chevron 〈 〉 denotes an average over patients. For each particular dose level D, patients were subsequently added to the cohort until the RMSE did not change significantly as the sample size further increased. That ensured the estimated prediction uncertainty be invariant under change of sample size; it also justified the sufficiency of the sample size of our study. The mean error, was also tracked as a measure of systematic error of our prediction.
The patient treated for bilateral disease was considered as two cases, one left, and one right. By minimizing RMSE for the first ten patients, it was determined that V(11 mm) best estimated ipsilateral V 20 Gy(RBE) . The sample size increased to 16 to achieve convergence of RMSE, as shown in Fig. 3  providing planning objectives to initiate the treatment planning process. Most of these methods start from a quantity related to distance-to-target or overlap volume.

| DISCUSSION
One of the strengths of our model is its simplicity. It can be applied to a new clinical case once the target contour is available and the IL is delineated. The only functions necessary are isotropic expansion, intersection, and volume calculation or voxel count.
Those functions are virtually available in any contouring module. An estimation may be obtained within minutes or even seconds. In contrast, developing a clinical treatment plan can easily take hours even after the contours are ready. If programming or scripting interface is available, the model could be easily implemented as a one-button solution.
F I G . 1. Treatment plan for a typical patient. The filled contour shown is the planning target volume.
ZENG ET AL.

| 55
Despite its simplicity, the model is effectively accounting for a number of patient-specific features in estimating dose-volume metrics. Since the prediction is derived directly from patient contours, the size of target and OARs (e.g., lung), the separation between target and OARs (e.g., thickness of rib cage), shapes, and/or curvatures are all implicitly considered in the model.
In our determination of the model parameter m, we increased the sample size until the convergence of prediction error was achieved (Fig. 3). That means further increasing the sample size in establishing the model would not improve the prediction accuracy.
In other words, the currently observed prediction error is the intrinsic limit of the model itself.  ing (e.g., heart, chest wall, rib). Those cases will inevitably become outliers for our model, one example of which is shown in Fig. 6. In that case, part of the PTV is underdosed in order to spare the heart, which may be a common compromise one has to make in breast cases where internal mammary nodes are involved. As shown in the V 20 Gy(RBE) was predicted to be 18%, while the actual value was 13%. V 5 Gy(RBE) was also overestimated to be 39%, while the actual value was 32%.
The dose predictions are limited by the accuracy of dose calculation algorithm, especially in lung. Pencil beam (PB) algorithm was used in this study. 16 The deficiencies of PB algorithm have been reported in the literature for many years. 17 Fig. 2), again less than the RMSE of our model prediction. As a result, our model can be directly applied to all the common prescription dose levels, including sequential boost, while acknowledging the RMSE of 5%-6%.
A major portion of the cohort (80%) had nodal involvement, corresponding to a variety of target shapes on which our model was validated. None of the cases was partial breast irradiation, for which lung dose is usually not a clinical concern.

| CONCLUSION
To conclude, we have established a contour-based model that enable prediction of IL V 20 Gy(RBE) and V 5 Gy(RBE) for breast cancer patients treated with proton therapy. The model is clinically easy to implement and provides an estimate of the expected values prior to planning.

CONF LICT OF I NTEREST
No conflicts of interest.