Can automated treatment plans gain traction in the clinic?

Abstract Recently, there has been an increased interest in the feasibility and impact of automation within the field of medical dosimetry. While there have been many commercialized solutions for automatic treatment planning, the use of an application programming interface to achieve complete plan generation for specific treatment sites is a process only recently available for certain commercial vendors. Automatic plan generation for 20 prostate patients was achieved via a stand‐alone automated planning script that accessed a knowledge‐based planning solution. Differences between the auto plans and clinically treated, baseline plans were analyzed and compared. The planning script successfully initialized a treatment plan, accessed the knowledge‐based planning model, optimized the plan, assessed for constraint compliance, and normalized the treatment plan for maximal coverage while meeting constraints. Compared to baseline plans, the auto‐generated plans showed significantly improved rectal sparing with similar coverage for targets and comparable doses to the remaining organs‐at‐risk. Utilization of a script, with its associated time saving and integrated process management, can quickly and automatically generate an acceptable clinical treatment plan for prostate cancer with either improved or similar results compared to a manually created plan.


| INTRODUCTION
Concerning prostate cancer, it has been stated "that all men, if they live long enough, can expect to get the disease." 1 Prostate cancer ranks as the second most common cancer for American males behind skin cancer, representing 20% of the cancer diagnoses for men among all specific disease sites 2 with surgery, radiotherapy, and active monitoring 3 as typical treatments. Creation of a radiotherapy treatment plan is a complex process that produces a unique result with the potential for disparity in plan quality achieved and planning time invested. Studies have shown plan quality can affect patient outcomes, with a definitive benefit to those techniques that allow dose escalation while limiting high rectal tissue dose to avoid toxicities. [4][5][6][7] Common external beam treatment modalities utilized for prostate cancer include three-dimensional conformal radiotherapy (3D-CRT), fixed gantry angle intensity-modulated radiotherapy (IMRT), as well as volumetric-modulated arc therapy (VMAT). 8 Both IMRT and VMAT are highly conformal, inversely planned approaches that have become the standard of practice for the treatment of prostate cancer. 9,10 VMAT in particular has experienced a quick and collective rise in part due to increased efficiency and speed of delivery with sustained and pronounced quantitative quality, potentially improved dose homogeneity, and enhanced normal tissue sparing compared to IMRT and 3D-CRT techniques. 8,11 Currently, the creation of VMAT plans is still an intricate and time-consuming process due to com- quality metrics and production times. 12 The net result is an iterative process of attempted dosimetric improvement, potentially incurring excessive time expenditure for minimal clinical return on investment.
One solution designed to reduce planning time while producing consistent, high quality, and clinically acceptable treatment plans is Knowledge-based planning (KBP). 13,14 By leveraging dosimetric and geometric information from previous clinical plans, KBP has the capability to reduce the plan variation, and to optimize time investment in a manner typically associated with increased treatment planner expertise. 14 While field design and inverse planning options can be implemented through ESAPI, each patient has unique anatomical relationships that require intricate digital analysis and complex logic, which causes scripting alone to be a suboptimal approach. Decision-making in the planning process is highly subjective and remains dependent on the knowledge, experience, and capability of the planner. 16,17 With the release of Eclipse version 15.5, ESAPI now has the ability to write to the ARIA database. This creates new possibilities to leverage the power automation and KBP simultaneously to move closer to the theoretical ideal plan state without user interaction.
Furthermore, stand-alone access to ESAPI has made it possible for plans to be created, initialized, and optimized outside the context of the current TPS user interface altogether. The purpose of this study is to demonstrate that with this combination of features we can obtain clinically preferred treatment plans for prostate cancer treatment with almost no user intervention. The AP script requires a CT data set and radiotherapy structure set objects to exist in the database. When initiated, it checks the integrity of the input data, creates the course, initializes a treatment plan with those basic elements, places the isocenter, enters the prescription according to the electronic records, generates standard two full arc beam arrangement, adds the reference points, and customizes the dose calculation settings. Course, plan, and beam names are automatically rectified from default values to match departmentspecific naming conventions.
Once initialized, the RapidPlan™ model is applied to the structure set and DVH estimates are generated. This model automatically 30 | generates the optimization parameters based on departmental specifications set by the user during initial model training and that patient's unique anatomy. The optimization process is started, dose calculated, and standard normalization applied. Next, the resulting plan is evaluated against key dosimetric endpoints and if they are not met then additional optimization points are added, and the plan is re-optimized. The final plan is then saved back to the database and the user is notified of its completion. APs were generated on all 20 test cases without any additional user input, interactive intervention, or post calculation adjustments. The completed plans were verified for minimum target and OAR compliance through a scripted plan checker. of the PTV volume divided by D RX ) were selected. The selected parameters were mean dose and V 75Gy , V 70Gy , and V 65Gy for rectum, and mean dose and V 75Gy V 70Gy , and V 60Gy for bladder, respectively.

2.C | Dose comparisons and clinical review
Statistical analysis was performed to compare the dosimetric differences between AP and MP. Paired Student's t test was used to evaluate the statistical significance of all the dose-volume parameters between the MP and AP. A P < 0.05 was considered statistically significant.
To yield a real-world assessment of the robustness of the AP process, actual patients from a consecutive time-period were utilized yielding a range of target and normal tissue volumes. The AP script utilized only the PTV and OARs structures without the need for any planning or helper structures.
Treatment plan complexity was evaluated by the use of modulation complexity score (MCS). This algorithm was originally defined to assess the plan complexity and deliverability for step-and-shoot IMRT 18 and was extended to apply to VMAT treatment (utilizing control points of the arc to replace the segments). 19

| RESULTS
A representative coronal, sagittal, and axial isodose distributions for the AP and MP are shown in Fig. 1. Qualitative inspection shows that the prescription dose (yellow line) is more conformal for AP compared to MP. The volume 50% of the prescription dose (red line) is comparable between the two plans with slightly more left and right splay for AP but less in the anterior and posterior directions.
Location of the intersection of the 50% isodose line and the rectum is closer to the target for AP, indicating greater sparing of the rectal volume. The mean HI was slightly higher for AP plans; however, the target minimum dose shows improvement, easily visualized by the jagged yellow prescription isodose line, which shows "holes" of lower than prescription dose throughout the target volume on the MP.
For all 20 patients, both AP and MP plans meet departmental guidelines for OAR sparing, as well as the minimum target coverage. Table 1 compares average values of selected dose-volume parameters between AP and MP for PTV, rectum and bladder, as well as Pvalues from paired Student's t test. All percent values are based normalized to prescription dose or total structure where applicable.
Maximum dose (D max ) referred to throughout this work was calculated as the maximum dose to 0.035 cc of a given structure.

3.A | PTV coverage, HI, and CI
The ratio of the AP and MP for D max , D 95 , D 98 , HI, and CI for each case is shown in Fig. 2

for the PTV. For D max and HI with values <1
indicating better performance for AP, 1 indicating parity, and greater than 1 indicting better performance of the MP. For D 95 , D 98, and CI, this pattern is reversed as an optimal plan maximizing these values.

3.B | OAR sparing
There was significant improved sparing of the rectum for the comparison points V 70Gy , V 65Gy , V 60Gy , and V 55Gy (P-values from Table 1).
The ratio of the AP and MP for D max , D mean , V 75Gy , V 70Gy , and V 65Gy for each case is shown in Fig. 3, again with values less than 1 indicating better performance for AP, 1 indicating equivalence, and greater than 1 indicting better performance of the MP.
The results for bladder are summarized in Fig. 4, indicating the ratio of the AP and MP for D max , D mean , V 75Gy , V 70Gy , and V 60Gy . No significant differences are found between AP and MP for all the dose-volume parameters for bladder in this study with the exception of the max dose (see Table 1).

3.D | Blind review
The results of the blind review are summarized in Fig. 5 Fig. 2 hovering near the parity line shows the similarity in PTV coverage between AP and MP approach.
The results of Table 1 show increased sparing at points of interest within the rectal volume. This sparing is statistically significant as shown by the analysis. The superior sparing of the rectum is also visually apparent in Fig. 3  Additionally, this progression not only replicates the process of planning but also provides a complete and AP solution, from origin and initialization through optimization, normalization, and review.
This increases the overall value of this process of automation, greatly reducing the time required from simulation to plan review by combination of these two powerful tools.
Even if the AP is not a perfect solution for all patients, it still provides an excellent starting point for further manual optimization and plan improvement. The processes of setting the isocenter, creation of fields with following appropriate naming conventions, and design of optimization goals all being performed without user input has intrinsic value to busy, high-performing departments. Additionally, the application of KBP constraints, initial normalization, and creation of some optimization structures through the AP script such as a PTV expansion to help with coverage can assist if further optimization is necessary. It should prove to be a simple process to verify optimization weighting, ease or increase constraints, and manage priorities to help achieve a more idealized isodose distribution.
The results of the blinded review were extremely compelling.
Given the overwhelming vote of confidence in the AP-produced outcome, there is little reason not to utilize this solution as a primary approach. This revelation has proven to be a key motivating factor for planners in our clinic to adopt KBP solution through AP for clinical use where manual implementation alone was not.
Radiation treatment planning for prostate cancer is a common, well-established treatment site with clear clinical guidelines and was therefore ideal for a proof of concept study into automation-and knowledge-based planning. While these plans may be seen as straightforward to experienced planners, the broad clinical application of KBP suggests that application of this technique may easily extend to other more challenging anatomical sites. The ability of the system to overcome complexities, which were formerly insurmountable boundaries for a solely computerized system, shows promise for further exploration.

| CONCLUSION
Treatment planning scripts can be a valuable tool in the creation of plans based on both improved efficacy and efficiency of treatment.
Automated plans using KBP were able to produce plans of clinical be the preferred process for initiation of treatment plans. Although this study was limited to prostate cancer treatment as a proof of concept, we expect that the design is extensible to other anatomical sites and we plan to explore that in future work. Neither resource alone, scripting, or KBP have proven to be a complete or independent solution for automated plan generation but together we believe they can become an invaluable clinical tool in today's demanding healthcare environment.

CONF LICTS OF INTEREST
The authors do not have any conflicts of interest to report.