Automated planning of whole breast irradiation using hybrid IMRT improves efficiency and quality

Abstract Purpose To develop an automated workflow for whole breast irradiation treatment planning using hybrid intensity modulated radiation therapy (IMRT) approach and to demonstrate that this workflow can improve planning quality and efficiency when compared to manual planning. Methods The auto planning framework was built based on scripting with MIM and Pinnacle systems. MIM workflows were developed to automatically segment normal structures and targets, identify landmarks for beam placement, select beam energies, and set beam configurations. Pinnacle scripts were generated from the MIM workflow to create hybrid IMRT plans automatically. Each hybrid IMRT plan included two prescriptions: a three‐dimensional (3D) prescription consisted of two open tangent beams, and an IMRT prescription consisted of two step‐and‐shoot IMRT beams. The 3D prescription delivered a full prescription dose to the maximum dose point, and the IMRT prescription was optimized to deliver a uniform dose to the entire breast while sparing dose to the normal structures. For 30 patients, the auto plans were compared with clinically accepted manual plans using the paired sample t‐test. Results The auto planning process took approximately 8 min to complete. The mean dice coefficients between auto‐segmentation and manual contours were 0.98, 0.94 and 0.88 for the lungs, heart, and PTVeval_Breast, respectively. The MUs of the auto plans was on average 13% higher than that of the manual plans. Auto planning improved plan quality significantly: percentage volume receiving 95% of the prescription dose (V95%) of the PTVeval_Breast increased from 91.5% to 93.2% (P = 0.001), V105% of the PTVeval_Breast decreased from 7.2% to 1.2% (P = 0.013), V20Gy of the ipsilateral lung decreased from 13.1% to 10.4% (P = 0.001) and mean heart dose for left‐sided breast patients decreased from 1.2 Gy to 0.9 Gy (P < 0.001). Conclusion An automated treatment planning process can make the planning process efficient with improved plan quality.


| INTRODUCTION
Whole breast irradiation (WBI) using tangential fields is an established technique for adjuvant radiation therapy as part of breastconserving treatment of early-stage breast cancer. 1 The planning goals of WBI include delivering a uniform dose to the breast tissue while minimizing dose to the lungs and the heart. Based on patient-specific anatomy, planners manually select tangent beam configuration (gantry angles and collimator angles) and beam energies, design the beam apertures (including segments), and set beam weights through a time-consuming trial and error approach.
Planning time was on average 39 min (range: 15-70 min) in a 20 patient study. 2 Several auto and semi-auto planning techniques have been developed to improve the efficiency of each step of WBI treatment planning. These steps included automatically segmenting the treatment targets and normal structures, [3][4][5][6] automatically selecting beam angles, 7 optimizing wedge angles for tangent beams, 8 optimizing segment shapes and weights, 3,9,10 and automatically creating an inverse planned intensity modulated radiation therapy (IMRT) plan. 4,11,12 Automation can reduce planning time with comparable or better plan quality. Zhao et al. proved that automatic beam angle selection reduced the volume of heart receiving 5 Gy and volume of ipsilateral lung receiving 10 Gy. 7 Mitchell et al. used auto-segmentation software to contour critical structures and scripting in the treatment planning system to generate beam segments from isodose lines and optimize segment weights. 3 4 This technique used radio-opaque markers placed at CT simulation to determine the beam geometry and generate whole breast volume for inverse IMRT optimization. For the 158 patients studied, the mean planning time was 6.8 min. Ninety-nine percent of the auto plans were deemed clinically acceptable, and 87% were deemed clinically improved or equal to manual plans. Purdie's method was applied clinically to over 1600 patients and was shown to reduce plan rejection rates. 8 We adapted Purdie's method to our clinical practice and improved the workflow to overcome some limitations of the original technique: (a) we allow flexibility in determining treatment boundaries and do not require wiring the patient in a specific way; (b) we allow the use of mixed energies and perform beam weight optimization automatically; (c) we automate and standardize the use of heart and lung blocks; and (d) we use a new hybrid IMRT technique which can maximize the weight of open beam to improve delivery efficiency and robustness. In this study, we described our auto planning workflow and compared plans created with this workflow with clinical plans for volume delineation, beam arrangement, planning parameters, plan quality and delivery robustness.

2.A | Patients
This institutional review board (IRB) approved retrospective study included 30 patients randomly selected from patients treated at our institution with tangential field-in-field WBI between November 2016 and November 2018. In this cohort, 10 patients were left breast cancer treated with deep inspiration breath hold (DIBH), 10 patients were left breast treated without DIBH, and ten patients were right breast cancer treated. The clinical plans for these patients used 6/10 MV photon with four standard fraction and 26 hypofractionated schemes. All patients were planned using Pinnacle treatment planning system, version 9.10 (Philips Healthcare, Fitchburg, WI) and treated on Truebeam Linear accelerator (Varian Medical Systems, Palo Alto, CA).

2.B | Auto planning workflow
For this study, all patients were auto planned using the same dataset, treatment machine, isocenter, and prescription dose as the corresponding clinical plans. Figure 1 shows the auto planning workflow. It consists of the following steps: a Auto-segmentation of volumes of interests and beam boundary points Computed tomography images were sent to the MIM system (MIM Software Inc., Cleveland, OH, USA). An assistant rule had been set up in MIM to automatically perform atlas-based segmentation for normal structures including bilateral lungs, heart, and spinal cord using an MIM workflow and a breast atlas, which was developed based on 20 patients outside of this study cohort.
Physicians reviewed the CT images to determine the extent of breast tissue to be treated based on wires placed at simulation and clinical judgment. They had the option to redefine the medial and lateral boundaries (using a contour named "box") and the superior and inferior boundaries (using a contour named "borders"). Another MIM workflow was developed to detect four boundary points: medial, lateral, superior and inferior based on the "box" and "borders" contours if they existed. Otherwise, the wires were automatically detected and use for boundary placement. In addition to the boundary points, two chest wall points, one at the chest wall & superior boundary and one at the chest wall & inferior boundary of the treatment region were placed by the MIM workflow. The boundary points and chest wall points were used to set up beams. The middle panel in Fig. 1(a) illustrates how to define the boundary and chest wall points from contours and CT.
Once beam boundaries were defined, the MIM workflow continued to segment the target volumes following the definitions of RTOG 1005 (Table 1). 13 b Beam placement and beam parameters optimization Two tangent beams were created automatically using a Matlab After beam geometry optimization, if more than 10 cc heart volume were exposed in the beam, a heart block was added to block all or part of the heart without blocking the lumpectomy cavity with a 5 mm margin. For ipsilateral lung and the normal tissue inferior to the ipsilateral lung, a lung block was added if it did not block any part of PTVeval_Breast with 1 cm margin. F I G . 1. Auto planning workflow. (a) Auto-segmentation: normal structures including bilateral lungs, heart, and spinal cord were contoured using atlas-based segmentation; the superior, inferior, lateral and medical boundary landmarks were identified from the "box" and "border" contours or the wires placed on the patient skin. Two chest wall contours were identified from the computed tomography; targets were segmented following RTOG 1005 recommendations. (b) Beam placement: two tangent beams were placed based on the boundary points and chest wall points. The gantry angle, collimator angle, and jaw positions were then optimized to maximize target coverage and minimize normal lung and heart volume in the beam. Beam energy was selected based on the maximal separation; (c) Hybrid IMRT: the automatic breast plan includes two prescriptions: a three-dimensional (3D) prescription with two static tangent beams and an intensity modulated radiation therapy (IMRT) prescription with two step-and-shoot tangent beams. The beam weightings of the 3D prescription were optimized using dose points selected uniformly inside PTVeval_breast. The 3D prescription delivers full prescription dose to the maximum dose. The IMRT prescription was optimized to deliver uniform dose to breast and reduce dose to lungs and heart.
A patient-specific pinnacle script was created at the end of the MIM workflow for automatic hybrid IMRT planning.
c Hybrid IMRT planning The hybrid IMRT plan included a 3D prescription and an IMRT prescription. The 3D prescription was associated with a pair of open medial and lateral beams (with heart and lung blocks if used), and the IMRT prescription was associated with two step-and-shoot beams using the same beam geometry as the 3D tangent beams.
The energy of the beams was defined based on the maximum tangent separation, same as our clinical practice. If the maximum separation of the tangent beams was less than 20 cm, 6 MV was used for both 3D and IMRT prescriptions; for separation great than 23.5 cm, 10 MV was used for both 3D and IMRT prescriptions; for separation between 20 and 23.5 cm, 6 MV was used for the 3D prescription, and 10 MV was used for the IMRT prescription. If the tumor bed volume was within 5 mm from the skin, 6X was used for the 3D prescription regardless of the tangent separation. The 3D prescription delivered a full prescription dose to the maximum dose point. The weights of the medial and lateral 3D beams were initially set to be equal and then optimized (beam weight optimization) based on dose points uniformly placed within the breast tissue. IMRT prescription was optimized to increase homogeneity and reduce dose to the heart and lungs. Table 2 lists the optimization criteria. The IMRT plan allowed a maximum of 20 segments (10 segments for breath hold plans to reduce treatment time). Minimum segment area was set to 10 cm 2 to force the use of large segments. The "use current jaw as max" feature was checked so that the IMRT segments did not exceed the edges of open beams.

2.C | Comparing auto and clinical plans
The auto-segmented contours were compared with the clinical contours using the Dice coefficient and mean Hausdorff distance. For a fair plan comparison, the auto contours were discarded, and the clinical contours were used for subsequent beam placement and hybrid IMRT planning.
To compare beam parameters between auto plans and clinical plans, we calculated the absolute differences of gantry and collimator angles, the difference in the posterior jaw (X1), the inferior jaw (Y1) and the superior jaw (Y2). The anterior jaw X2 covered the entire breast with >=2 cm margin in both clinical and auto plans. The auto placed beams were used for hybrid IMRT planning.
To compare the plan quality between auto plans (with clinician derived volumes) and clinical plans, target coverage PTVeval_Breast V95 and PTVeval_Lumpectomy V95, high dose volume PTVeval_-Breast V105, ipsilateral lung receiving 20 and 5 Gy and heart mean dose were compared. Statistical analysis was performed using twosided, paired sample t-test in Excel with significance defined as P < 0.05.

| RESULTS
The entire auto planning workflow took approximately 8 min, of which auto-segmentation took about 3 min (without manual editing), beam placement took about 2 min, and hybrid IMRT planning took about 3 min.

3.A | Auto-segmentation
The auto-segmentation achieved good agreement with the clinical contours. Figure 2 shows an example. Table 3 lists the dice coefficients and mean Hausdorff distances between auto contours and clinical contours. The Dice coefficients were more than 0.9, and the mean Hausdorff distances were within 1.5 mm for all normal structures. No manual editing was necessary for the contours of the lungs or spinal cord while only small edits were needed for the heart contours. For targets, dice coefficients of 0.84 and 0.88 were achieved with CTV_breast and PTVeval_breast, respectively. Heart Max EUD (α = 1) < 1 Gy 1 For all patients, Fig. 4    ume. 4 We arbitrarily chose 20 as the maximum number of segments for free breathing patients, and 10 for breath hold patients. To study the impact of segment number, case #1 was planned with 5, 10, 20, and 50 maximum number of segments respectively. To further evaluate the impact of the auto-contouring errors on automatic breast treatment planning, for an example patient (case #1), Table 7 compares the plan parameters, dose distribution, and plan quality between the auto plan using auto contours and the auto plan using clinical contours. For this particular patient, the Dice coefficients between auto contours and clinical contours were 0.80, 0.98, and 0.92 for CTV_breast, lungs, and heart, respectively.

3.B | Auto plan VS clinical plan
Although target and normal structure volumes varied, the plan parameters (beam gantry and collimator angles, jaw sizes, and plan MU) were similar between the two plans. The 90% and 50% isodose volumes of the two plans overlapped by 94% and 95% respectively.
DVH metrics to targets and normal structures (evaluated using clinical contours) were comparable between the two plans. The auto planning technique introduced in this study is robust to the errors of auto-contouring. Two main reasons may explain the insensitivity of the auto plan to contouring errors: (a) the auto-contoured normal structures closely matched the clinically approved manual contours (>0.9 dice coefficients); (b) we applied restrictions in beam geometry optimization: beam gantry and collimator angles cannot deviate from initial angles by more than 10°and beams were not allowed to cross midline.
It is worth noting that although the plan parameters and dose distribution were robust to contour variations, DVH metrics were sensitive to how the targets and organs at risk were contoured. As shown in Table 7, PTVeval_Breast V95 from the auto plan was lower when evaluated with the auto contour than with the clinical contour.
As shown in Fig. 2, the auto-segmented CTV_Breast and PTVeval_-Breast was larger than the manual contours. The heart mean dose was slightly lower when evaluated with the auto contour than with the clinical contour, again due to smaller auto-contoured heart than the clinical contour. Therefore, it is still recommended to review and edit the auto contours for plan evaluation.

| CONCLUSION S
An automated treatment planning technique was developed for whole breast irradiation using hybrid IMRT. Compared with manual planning, auto planning improved planning efficiency and plan quality. A future study will focus on the assessment of the robustness of auto plans with more patient data.