Implement a knowledge‐based automated dose volume histogram prediction module in Pinnacle3 treatment planning system for plan quality assurance and guidance

Abstract Purpose This work aims to develop a knowledge‐based automated dose volume histogram (DVH) prediction module that serves as a plan quality evaluation tool and treatment planning guidance in commercial Pinnacle3 treatment planning system (Philips Radiation Oncology Systems, Fitchburg, WI, USA). Methods The knowledge‐based automated DVH prediction module was developed with kernel density estimation (KDE) method and applied for Pinnacle3 treatment planning system. Treatment plan data from 20 esophageal cancer cases were used for creating a module to predict DVHs. Twenty additional esophageal clinical plans were evaluated on the developed module. Predicted DVHs were compared with manual ones. Differences between the predicted and achieved DVHs were analyzed. Results The plan evaluation module was successfully implemented in Pinnacle3 treatment planning system. Strong linear correlations were found between predicted and achieved DVH for organs at risk. Suboptimal treatment plan quality could be improved according to the predicted DVHs by the module. Conclusion The knowledge‐based automated DVH prediction module implemented in Pinnacle3 could be used to efficiently evaluate the treatment plan quality and as guidance for further plan optimization.


| INTRODUCTION
Intensity-modulated radiation therapy (IMRT) is a popular clinical treatment modality used worldwide. Compared to conventional beams of uniform intensity, intensities of radiation beams are modulated in IMRT to deliver a nonuniform dose distribution to the tumor target. A desired target dose conformity and sufficient sparing of critical structures could be achieved through IMRT planning. 1-3 An ideal IMRT plan could require a lot of trial-and-error process and was time-consuming. The efficiency of making an IMRT plan and the quality of this plan depend on the clinical experience of the dosimetrist. [3][4][5] The population of patients treated for esophageal cancer is increasing in our center. Esophageal plan quality has been investigated for different treatment modalities. 6 Compared to three-dimensional conformal radiotherapy (3DCRT), IMRT plans generally show better target dose coverages and lower mean doses to organs at risk (OARs). Also, high-quality plans are desired to achieve optimal treatment outcomes. 7 Therefore, a pretreatment plan quality assurance (QA) tool is crucial for assuring treatment plans.
Knowledge-based dose volume histogram (DVH) prediction has shown good results for head and neck, pancreas and prostate treatment planning. [8][9][10][11][12] The plan variability can be reduced through knowledge-based planning. [13][14][15][16] Although there is no definitive method to evaluate a radiotherapy treatment plan, the feasibility of using knowledge-based DVH prediction for QA of head and neck plan has been demonstrated. 17 As a commercial knowledge-based optimization engine, RapidPlan™ in the Eclipse treatment planning system (Varian Medical Systems, Palo Alto, USA) showed its potential to serve as an accurate plan QA tool. Strong correlations had been found between RapidPlan™ predicted and manually achieved mean doses to OARs. The predicted OAR sparing was validated by replanning the patient with the RapidPlan™ module. Suboptimal IMRT plan quality can be improved. Besides the knowledge-based RapidPlan™ module in the Eclipse, Pinnacle 3 treatment planning system (Philips Radiation Oncology Systems, Fitchburg, WI, USA) also provides an Auto-Planning module using plan-simulating for automated plan optimization. 15,18 However, it cannot predict OAR DVH.
The dilemma on balancing plan efficiency and plan quality is popular in developing country with heavy patient load but limited dosimetrists. Therefore, a quick method to verify plan quality or guide the treatment planning is crucial in clinic.
We have proposed a new knowledge-based auto-planning solution for IMRT treatment planning. 19 It used two parameters kernel density estimation (KDE) algorithm to build the DVH prediction model. The model feasibility in breast cancer has been validated with treatment plan quality and consistency. In this study, a different KDE method was developed and applied for esophageal cancer and integrated into the commercial Pinnacle 3 treatment planning system to create a knowledge-based automated DVH prediction module, which can be used for treatment planning QA or guidance.
2 | ME TH ODS AND MATERIALS 2.A | Overview of the knowledge-based plan QA module workflow A schematic view of the plan QA module in Pinnacle 3 is showed in Fig. 1. Twenty IMRT cases with the same prescription dose for esophageal cancer were planned by dosimetrists who had more than 10 years work experience in our institution. These manually optimized plans were clinically acceptable with high quality. DVH data, voxels of planning target volume (PTV) and OARs extracted from these plans were used as training data. A KDE model was trained with these data and integrated into the Pinnacle 3 treatment planning system as a plan evaluation module. For new patient with PTV and OAR delineation, corresponded DVH prediction would be generated from this module. Predicted DVH curves can be shown on Pinnacle 3 user interface (UI), which could be used as a benchmark to evaluate treatment plan quality or as guidance for further treatment optimization.

2.B | KDE-based DVH prediction
Dose distribution, voxels of PTV and OARs extracted from the training data were trained to get DVH prediction. DVH can be determined by a cumulative probability distribution of dose x that is lower or equal to a given dose D as Eq. (1).
The dose probability density function p D (x) was predicted by marginalizing the conditional probability density p(x|t) estimated from training data over probability density p*(t) from input plan data, where t was the signed minimal distance between dose point in OAR voxel and PTV boundary. When OAR overlapped PTV, signed distance of the point was negative inside PTV. For each input plan data, p D (x) can be estimated by Eq. (2).
1. Schematic view of knowledge-based plan quality assurance module workflow.

2.D | Plan QA module implementation and user interface integration in Pinnacle 3
Realization of plan QA function in Pinnacle 3 treatment planning system has three parts: plan data access, data processing, and results feedback. Plan data from treatment planning system was imported through an interface developed in Java programming language and data processing was also performed in Java. The plan QA model was developed in R programming language. The predicted DVH from the model was embedded in the plan evaluation module of treatment planning system using Pinnacle 3 's scripts.

2.E | Application and evaluation in clinical implementation
The developed QA module in Pinnacle 3 was applied for 20 new esophageal IMRT plans with the same prescription dose. Plan quality evaluation was made with respect to the DVH metrics in Table 1.
The discrepancy between predicted and achieved DVH metrics was analyzed statistically.

2.F | Statistical analysis
Clinical IMRT plans of esophageal cancer were evaluated between the manual DVHs and predicted DVHs from the QA module. Mean values and RMSE of data were used for statistical analysis. The statistics were calculated using SPSS (v13.0, IBM corp., New York, NY, USA). Regression analysis used R-squared (R 2 ) as the coefficient of determination. The coefficient of determination of the regression model would be close to 1 for a good fit.

3.A | KDE model validation
Performance of the KDE model for esophageal cancer plan is shown in Table 2. The mean values of the differences between predicted and achieved OAR metrics from validation plans were also summarized in the table. A 3.4% deviation of V 5 index of lung-PTV was found on average.

3.B | Pinnacle 3 user interface
The developed Pinnacle 3 UI for plan QA module has the following three parts: data base server login [ Fig. 2(a)]; patient plan selection [ Fig. 2(b)]; PTV and OAR matching [ Fig. 2(c)]. In this study, 20 patients were selected using this UI for evaluation.
Predicted DVH for plan QA was successfully integrated in the DVH display module in Pinnacle 3 using the script. An example of plan evaluation UI in Pinnacle 3 is shown in Fig. 3. The dashed line represents the predicted DVH and the solid line is the achieved DVH for the same OAR in treatment plan.

3.C | Clinical implementation
The dose comparison between predicted and manual DVHs was performed on 20 clinical plans. The OAR DVH comparison is shown in Table 3. A 1.2% deviation of V 20 index and 2.5% deviation of V 5 index of lung-PTV were found on average.

3.D | Predicted DVH-guided treatment planning
To test the feasibility of treatment planning guidance using the pre- With the guidance of predicted DVH, a re-optimization was performed. The final plan with re-optimization is shown in Fig. 6. This indicated the feasibility of using predicted DVH as a treatment planning guidance.

| DISCUSSION
The scope of this study was to apply a knowledge-based automated DVH prediction module that could be integrated in Pinnacle 3 UI for plan QA and guidance. To the best of our knowledge, this is the first time to develop the KDE-based DVH prediction module in Pinnacle 3 .
One of the 20 clinical esophageal IMRT plans was randomly selected out for re-optimization under the guidance of predicted OAR DVHs.
It was available to achieve an optimized result. We believe that this tool would be helpful for both plan QA and further DVH-guided plan optimization or re-optimization.