The benefits evaluation of abdominal deep inspiration breath hold based on knowledge‐based radiotherapy treatment planning for left‐sided breast cancer

Abstract Purpose To study the impact of abdominal deep inspiration breath hold (DIBH) technique on knowledge‐based radiotherapy treatment planning for left‐sided breast cancer to guide the application of DIBH technology. Materials and methods Two kernel density estimation (KDE) models were developed based on 40 left‐sided breast cancer patients with two CT acquisitions of free breathing (FB‐CT) and DIBH (DIBH‐CT). Each KDE model was used to predict dose volume histograms (DVHs) based on DIBH‐CT and FB‐CT for another 10 new patients similar to our training datasets. The predicted DVHs were taken as a substitute for dose constraints and objective functions in the Eclipse treatment planning system, with the same requirements for the planning target volume (PTV). The mean doses to the heart, the left anterior descending coronary artery (LADCA) and the ipsilateral lung were evaluated and compared using the T‐test among clinical plans, KDE predictions, and KDE plans. Results Our study demonstrated that the KDE model can generate deliverable simulations equivalent to clinically applicable plans. The T‐test was applied to test the consistency hypothesis on another ten left‐sided breast cancer patients. In cases of the same breathing status, there was no statistically significant difference between the predicted and the clinical plans for all clinically relevant DVH indices (P > 0.05), and all predicted DVHs can be transferred into deliverable plans. For DIBH‐CT images, significant differences were observed between FB model predictions and clinical plans (P < 0.05). DIBH model prediction cannot be optimized to a deliverable plan based on FB‐CT, with a counsel of perfection. Conclusion KDE models can predict DVHs well for the same breathing conditions but degrade with different breathing conditions. The benefits of DIBH for a given patient can be evaluated with a quick comparison of prediction results of the two models before treatment planning.


| BACKGROUND
Postoperative adjuvant radiotherapy (RT) plays an indispensable role in breast-conserving treatment to minimize the risks of local-regional recurrence and metastasis. Whole breast irradiation (WBI) after breast-conserving surgery as a comprehensive treatment model, which has been confirmed to possess the similar local control and overall survival rates to modified radical surgery in breast cancer patients. 1 However, the dose of the surrounding critical organs-atrisks (OARs), especially the heart, left lung and the left anterior descending coronary artery (LADCA), [2][3][4] are crucial to the RT quality assessment for left-sided breast cancer.
Therefore, by using a diversity of methods, such as DIBH, intensity modulated radiation therapy (IMRT) techniques, treatment in the prone position and proton therapy, to shield the heart and minimize the lung and LADCA doses while ensuring enough dose in the target volume during left-breast postoperative radiotherapy have been presented. Comparing DIBH and IMRT, IMRT is the most commonly used strategy in left-sided breast postoperative radiotherapy. [5][6][7] The DIBH maneuver we used is the abdominal DIBH (A-DIBH), which could widen the spatial Euclidean distances between the heart and the target volume. IMRT treatment technique has the capability of reducing the cardiac dose while delivering adequate target coverage because of its unique dose calculation and beam weight optimization.
The selection of the final radiotherapy regimen (especially the selection of respiratory mode) will greatly affect the normal tissue complications (NTCP) and tumor control rate. 8,9 In the previous IMRT plans, physicians often determined the ideal OAR dose volume limit through population-based recommendations (either from the tumor radiotherapy team or from the doctor's intuition). 10 However, the huge geometric differences in the complexity of PTV and OAR among patients make it a challenge for doctors to quickly and accurately select the best ultimate treatment for a particular patient within all acceptable options.
Knowledge-based planning (KBP) is a promising technology.
There is a large amount of image information and dose planning information of cancer patients in the current radiotherapy system, which has become a priori knowledge. By feature extraction and quantitative analysis of these prior knowledge, a reliable empirical model (KBMs) can be obtained to realize the intelligence of the radiotherapy planning system. Current studies have proved that KBP has a higher consistency of plan quality and higher operational efficiency than manual plans with different quality. [11][12][13] For example, RapidPlan™(Varian Medical Systems, Palo Alto, CA, USA) has been widely used as a commercial KBP product. 11,14 In the KBP method, the prediction of DVH in new patients requires the use of the DVH of OAR in the previous clinical plan and the parameterized model generated by the relevant anatomical structure, 13,15 thus emphasizing the importance of the parameterized prior model. However, it remains to be seen whether the implementation effect of the parameterized prior model in KBP is consistent under different breathing conditions. To our best knowledge, the impacts of different breathing methods during CT simulation for leftsided breast cancer on knowledge-based treatment planning have not been reported before. Therefore, this study established two knowledge-based empirical models for the treatment of the same group of breast cancer patients based on different respiratory conditions. We then used these two KBMs to cross-predict CT in both breathing patterns, creating four KBP plans for each patient. We attempted to investigate the compatibility of KBP with different respiratory conditions, such as whether the DIBH KBM is applicable to FB-CT prediction, or whether the FB KBM is applicable to DIBH-CT prediction. Quantifying the benefits of using the specified model can help us clearly understand the use of KBM to predict the OAR dose of postoperative radiotherapy for breast cancer and guide the application of A-DIBH radiotherapy technology.

| METHOD AND MATERIALS
The workflow of this study is illustrated in the diagram in Fig. 1. DVHs of these ten new patients were estimated by two KBMs.
The estimated DVHs were taken as a substitute to dose constraints and objective functions in Eclipse treatment planning system (TPS) for each new patient.
There are three types of comparisons we want to investigate in this study. Firstly, we wanted to confirm whether our model could precisely predict the DVHs with same breath settings, such as using FB model to predict patients with FB-CT. These comparisons were marked "green" in Fig. 1.
Secondly, we wanted to investigate whether the model built with one breath condition can precisely predict DVH for patients with another breath. Such as using FB model to predict the DVH for the patient with DIBH-CT images. These comparisons were marked "blue" in Fig. 1.
Third, we wanted to investigate whether these DVH prediction models can be optimized to a deliverable plan. These comparisons were marked "red" in Fig. 1. To achieve optimal homogeneity of the data in the present analysis, we incorporated only the whole-breast irradiation series. Target volumes and OAR were entirely contoured via two CT series in the Eclipse treatment planning system (Varian Medical Systems, Palo Alto, CA, USA) according to the Danish Breast Cancer Cooperative Group (DBCG) atlas. 16 Two intensity modulated radiotherapy treatment plans were generated in the Eclipse for each CT, using the anisotropic analytical algorithm (AAA) for final dose calculation. All IMRT plans containing 6 fixed non-opposing fields, and thegantry angles and beam energies of each plan are the same as the clinical methods.

2.A | Patients and treatment planning
The criterion of treatment plans was that 97% of the PTV should be covered by at least 95% of the isodose (and <108% of the isodose), and the mean dose of PTV in the whole cases was prescribed to 50 Gy in 25 fractions.

2.B | KDE model training
Inspired by Skarpman's KDE algorithm, 15 the two-parameters KDE which incorporated two predictive features was implemented to predict DVHs. It calculates the conditional probability density of the dose d given x (the signed minimal distance between the voxels on the PTV surface and in the OAR) and θ (the angle between x and the center of the CT image) from the training dataset.

2.D | Dosimetric comparison
A paired student's T-test was used to assess the significance of any differences in dose metrics where significance corresponded to a P-

3.A | The performance of the KDE models
The results (Table 2) show that there were no differences between clinical plans and KDE predictions for both models in the training dataset, confirming that the DVH estimation of the KBM was successful.

3.B | The models work in same breath settings
The results of model performance in the same breath settings for another 10 left-breast patients are presented in Table 3

3.C | The FB model works with DIBH-CT
The result of the FB model works with DIBH-CT was presented in  Table 5 shows the result of the DIBH model works with FB-CT.

3.D | The DIBH model works with FB-CT
Compared to the clinical manual plans, the KDE prediction resulted in lower mean doses of the heart and LADCA by 0.24 AE 0.36 Gy (P = 0.02), and 4.57 AE 2.46 Gy (P = 0.014), respectively. The left lung mean dose of the KDE prediction was 0.33 AE 0.99 Gy higher than the clinical plan (P = 0.01).
Significant differences were observed in all structures between the KDE plan and KDE prediction (P < 0.05 for all three structures).
These predicted DVHs may not be directly transferred to a deliverable plan.

| DISCUSSION
Deep inspiration breath hold offers increased lung volume and suppressed respiratory motion. As Schönecker et al. 17     We consider this study innovative because it explains and validates the correlation between classification of the KDE-based dose prediction model and breathing maneuvers during left-sided wholebreast irradiation after breast-conserving surgery. It shows that classifying KDE dose prediction models according to respiratory patterns are indispensable, thus an optimal decision base for automatically making the radiation therapy plan of the marshalling station with computer is supplied. This research, while just a beginning, at least establishes some basic scientific facts that could prove useful in future studies on the automatic plan and related conditions.

| CONCLUSION
The KDE model predicted DVHs and auto-plan DVHs were not significantly different from the clinical DVHs when the appropriate model is used (FB model for FB plan, DIBH model for DIBH plan).
The DIBH model should not be used for predicting FB treatments and vice versa. In view of the effective of the established model, the benefits of DIBH for a given patient can be quickly assessed before planning.

CONFLI CT OF INTEREST
None.