Single click automated breast planning with iterative optimization

Abstract Purpose To present the development of an in‐house coded solution for treatment planning of tangential breast radiotherapy that creates single click plans by emulating the iterative optimization process of human dosimetrists. Method One hundred clinical breast cancer patients were retrospectively planned with an automated planning (AP) code incorporating the hybrid intensity‐modulated radiotherapy (IMRT) approach. The code automates all planning processes including plan generation, beam generation, gantry and collimator angle determination, open segments and dynamic IMRT fluence and calculations. Thirty‐nine dose volume histogram (DVH) metrics taken from three international recommendations were compared between the automated and clinical plans (CP), along with median interquartile analysis of the DVH distributions. Total planning time and delivery QA were also compared between the plan sets. Results Of the 39 planning metrics analyzed 23 showed no significant difference between clinical and automated planning techniques. Of the 16 metrics with statistically significant variations, 2 were improved in the clinical plans in comparison to 14 improved in the AP plans. Automated plans produced a greater number of ideal plans against international guidelines as per EviQ (AP:77%, CP:68%), RTOG 1005 (AP:80%, CP:71%), and London Cancer references (AP:80%, CP:75%). Delivery QA results for both techniques were equivalent. Automated planning techniques resulted in an average reduction in planning time from 23 to 5 minutes. Conclusion We have introduced an automated planning code with iterative optimization that produces equivalent quality plans to manual clinical planning. The resultant change in workflow results in a reduction in treatment planning times.

has a significant role to play in the treatment of breast cancer; of the 16,000 incidences of breast cancer in Australia in 2015, a total of 13,969 (approximately 87%) cases received radiotherapy treatment. 3 With such a large number of clinical cases, breast radiotherapy shows the potential for significant efficiency gains with the introduction of automation. This work focuses on the potential reduction in planning times and changes in the plan-approval workflow in breast radiotherapy.
Previous papers in the literature have described methodologies for automated planning of breast patients using heuristic optimization, 4,5 external markers, 4-6 database libraries, [7][8][9] semiautomatic optimization by component protocol scripts, 9,10 and automatic optimization following manual beam setup. 11 . Some initial papers are beginning to show promising results for machine learning-based techniques for automated planning of breast patients. 12 We instead propose a fully automated method (auto-plan) that mimics the process adopted by human dosimetrists in iteratively optimizing with changing optimization weights dependent on plan performance against clinical objectives. To the author's knowledge this is the first presentation in the literature of a fully automated methodology that does not require a database of reference plans, nor hard-coded heuristic optimization objectives; instead, planning objectives are optimized iteratively in response to patient geometry/planning challenges.
The paper investigates the validity of such an approach for fully automated lumpectomy breast patient plans, including comparisons of plan time, consistency, quality, and deliverability with those of manually created plans that were previously accepted clinically.

2.A | Patient cohort
All comparisons in this paper were performed retrospectively on a cohort of clinical patients planned with tangential hybrid breast radiotherapy at the site between June 2016 and November 2017.
Patients were excluded from the study where nodal irradiation was included, or where a mastectomy had been performed. A total of 100 patients were selected for the study. All patients within the cohort were of stage 0, I, II breast ductal carcinoma in situ. All patient data used in this study were anonymized and all analysis and data mining were performed under the approval of the local institutional ethics board.
At the authors department two dose and fraction regimes are permitted for breast radiotherapy, namely 5000cGy in 25 fractions and 4240cGy in 16 fractions. The distribution of patients in this study was 57 right breast (40 at 5000cGy, 17 at 4240cGy) and 43 left breast (32 at 5000cGy, 11 at 4240cGy) patients. All patients within the study had received and completed whole breast hybrid IMRT tangent radiotherapy.

2.B | Clinical planning
The hybrid technique has been shown to provide high-quality plans that are robust to patient breathing motion. [13][14][15][16] The primary characteristic of the technique is conformal fields that deliver the majority of breast dose accompanied with a small weighting of modulated multileaf collimator (MLC) fields to fine tune the distribution.
Clinical plans (CPs) in this study applied the hybrid technique using an extensive clinical protocol with asymmetric conformal tangential fields and dynamic MLC (DMLC) modulated tuning fields.

2.C | Automated Planning
The automated plans (APs) were created using code written in Iron-Python 2.7 and run through the Raystation 6.0 planning system (Raysearch Laboratories, Stockholm, Sweden) scripting environment.
The entire code implementation is the combination of several internally developed modules, which are created to run on the standard Raystation IMRT license without installing further python libraries and the need of previous plan databases.
The AP code is reliant on approved contours ( Figure 1) from the oncologist/dosimetrist for standard breast planning. These are outlined in guidelines set out by ESTRO 18  These contours are generated by auto-segmentation code not explored in this paper. The auto-generation is a combination of atlas-based segmentation, heuristic ROI algebra, and intensity-based structure optimization. Also created automatically are relevant tattoo markers, a tangent geometry, medial and posterior border points, and planning structures (including rings and dose control volumes).
All auto-segmented structures are reviewed and edited as necessary by Oncologists prior to planning.
On initiation of the AP code, a user interface is presented. Users are required to specify target dose, fractionation, treatment machine, delivery technique, the primary PTV, and any boost PTV or nodal PTV to be treated. For the purpose of this paper the breast PTV was selected to be the only PTV and the Hybrid technique was selected as the only planning technique. All CPs and APs were created with a Varian iX beam model. The machine is dual energy (6MV and 18MV), with a millennium 120 leaf MLC of 0.5cm leaf width centrally (inner 80 leaves) and 1cm leaf width at larger field sizes (outer 40 leaves). The machine has a maximum possible field size of 40cm x 40cm with leaf modulation possible along all MLC pairs. The automated planning functionality performed by this code includes;

2.D | Automated beam generation
The tangent beam geometry is created by the auto-segmentation code. The position is optimized by incremental translation and affine transformations such that the intersection of the tangent with the external contours approaches the intersection of the tangent and PTV plus a volume margin to account for the field edge-to-PTV margin. The optimization of the tangent geometry is simplified by a condition of entry through the midline tattoo as per the protocol of the center which primarily limits the transformations to rotations pinned at the midline tattoo. An optimized tangent geometry is shown in Figure 1. The medial and posterior border are calculated from the intersection of the posterior aspect of the tangent with the tissue edge at the superior/inferior mid-plane level of the PTV. This is performed through a series of code driven algebraic functions that determine the outer rim of the tangent geometry and external contours and take their intersections at the level of the midline tattoo.
Simple trigonometry (equations 1 and 2) is utilized to determine the optimal medial gantry angle for the tangent beams, rounded to the nearest full angle as per clinical protocols at the site. Lateral angles are taken at the opposing angle to promote a closed-jaw nondivergent posterior edge.
For left-sided breast patients; For right-sided breast patients; where; • (x m , y m ) are the x and y (lateral and anterior-posterior) DICOM coordinates of the medial border, • (x p , y p ) are the DICOM coordinates of the posterior border; and • θ m is the medial and posterior tangent gantry angles. x iso , y iso ¼ x m AE10Þ, y m AE 10 Â y m À y p x m À x p , x m À x p j j≥10 where; • x iso , y iso are the x and y co-ordinates of the plan isocenter • AE is dependent on laterality of patient  hybrid IMRT technique. [14][15][16]20,21 At this point the plan is passed to the iterative optimization code.

2.E | Iterative optimizer
The iterative optimizer is code developed by the authors within While other groups have provided heuristic-and knowledgebased optimization, [4][5][6]8,9 where optimal objective functions are predicted, the iterative nature of this code provides a method potentially more able to deal with large changes in patient geometry.
Small patient changes that would result in poor plan quality with initial objective functions are tailored by the incremental adjustment of weightings, in the same manner that human dosimetrist would adjust IMRT weightings to improve plan quality. This also means that this iterative optimization code can be applied to any anatomical sites and delivery techniques without feeding it with additional rules or training sets, a potential advantage over machine learning or regression techniques.
All AP plans were created blind of the associated CPs, with no input of the quality of the retrospective CPs. All plans were free of any human interactions with the exception of starting the script. The automated plans were created on the identical protocol as the CPs and variations in beam geometry, weighting, and optimization are purely a result of variations between human and software interpretation of optimal solutions.
Following the iterative optimizer all plan DVH statistics were exported for analysis within the AP code.

2.F | Plan quality
Analysis was performed separately depending on the laterality of the disease (left or right). While two dose regimes exist and were used in planning, for simplicity of presentation all results were scaled to 5000cGy equivalent for evaluation. As an example, for a plan the Australian EviQ guidelines, 23 and the London Cancer breast radiotherapy guidelines. 24,25 A summary of the constraints used for analysis are shown in Table 1.
A comparison of clinical and automated beam angles was performed across the 100-patient cohort to assess the ability of the automated software to correctly identify the appropriate angles for a given patient geometry.

2.G | Deliverability
To ensure the created auto-plans could be delivered on treatment, a random selection of 10 auto-plans was delivered to an ArcCheck quality assurance (QA) device (Sun Nuclear Corporation, Florida, USA). Gamma analysis was performed at a global 3%/3mm criteria with a 10% threshold, the result of which was compared to that of the QA of the corresponding clinical plan.

3.A | Plan quality
The dose volume histograms for both the CP and the AP plans of both left and right breast patients are shown in Figure 3. Comparisons of the evaluation metrics for each dose regime and laterality are shown in Table 2. P values and significance in variance are displayed for each metric along with the dosimetrically superior planning technique.
Of a total of 39 metrics analyzed, 23 showed no significant variation between CPs and APs. Of the 16 metrics that did show statistically significant variations, 14 were improved on the APs.
The two metrics in which the CPs were superior were d95 PTV coverage in right breast patients, and the mean dose of the contralateral breast in left-sided patients. The average improvements in these metrics were 22cGy and 2cGy, respectively.
In comparison against national/international breast protocols the AP plans adhered to 77%, 80%, and 80% of cases for ideal constraints of RTOG, EviQ, and London Cancer guidelines, respectively. Of the 100 patients analyzed, the variation in automated and planned beam angle was within 1 degree in 72% of cases (0°-34, 1°-38) with an overall mean variation of 0.17 +/-1.97 degrees.

3.C | Planning time
either planning technique, although APs tended to produce lower heart doses. It should be emphasized that the improvements are statistically significant but not necessarily clinically significant in all cases.
The trends in the statistics are supported by lower heart and lung doses on the DVH distributions in Figure 3. Another advantage of the automated planning implementation is an improved consistency in plan quality, with reduced first and third quartile ranges and standard deviations for the contralateral breast, ipsilateral lung, and heart distributions in comparison to clinical plans.
The contralateral breast mean dose metric was the only statistically significant improvement in the CPs for left-sided patients. Conversely there appears to be a trend toward slightly higher contralateral breast doses in right-sided CPs, although the difference is not statistically significant. Considering this effect is not mirrored in right-sided cases and the slightly lower (although not significant) heart doses in the AP cases, this suggests a divergence of the relative importance given to heart and contralateral breast between human and software optimization, even in the presence of identical clinical goals.

4.A | Deliverability
Automated and clinical plans show equivalent deliverability and QA results. Little can be drawn from the slight 1.1% improvement in average passing gamma points of the APs, with no statistical significance (P = 0.0714) owing to a small sample size. While a greater number of QA deliveries may result in statistical significance, as the primary concern of this paper is plan quality and efficiency gains, the current analysis is sufficient to show equivalence in deliverability and further QA measurements were not undertaken.

4.B | Planning time
The implementation of automated planning shows considerable advantages in both efficiency and workflow processes over manual clinical planning. The results in Table 3   It should be noted that the current implementation of automated planning is not inclusive of auto-contouring, which is accommodated by a separate script in the clinic. While as previously discussed, this results in a small time cost in comparison to some AP techniques that integrate auto-contouring, 7,8 ensuring oncologists review contours prior to planning allows for flexibility of implementation with various techniques. By determining optimal tangent angles by adherence to approved oncologist contours, the code has the potential to support step and shoot IMRT, short-arc breast VMAT, breast and node VMAT, and integrated boosts delivery techniques. The variability of nodal treatment extent poses a difficult problem for automated planning in the absence of volume review prior to plan creation.
Following the results of this research the described code has been implemented clinically at the authors' site for both hybrid DMLC and step and shoot whole breast techniques.

| CONCLUSION
The IronPython programming language has been used to develop an iterative optimization platform that automates the planning process of inversely planned radiotherapy. In application of the code on hybrid breast patients the automated plans have shown to be equivalent to clinically accepted plans across 100 retrospectively analyzed patients. While