Machine learning-based automatic proton therapy planning: Impact of post-processing and dose-mimicking in plan robustness.

Automated treatment planning strategies are being widely implemented in clinical routines to reduce inter-planner variability, speed up the optimization process, and improve plan quality. This study aims to evaluate the feasibility and quality of intensity-modulated proton therapy (IMPT) plans generated with four different knowledge-based planning (KBP) pipelines fully integrated into a commercial treatment planning system (TPS). A data set containing 60 oropharyngeal cancer patients was split into 11 folds, each containing 47 patients for training, five patients for validation, and five patients for testing. A dose prediction model was trained on each of the folds, resulting in a total of 11 models. Three patients were left out in order to assess if the differences introduced between models were significant. From voxel-based dose predictions, we analyze the two steps that follow the dose prediction: post-processing of the predicted dose and dose mimicking (DM). We focused on the effect of post-processing (PP) or no post-processing (NPP) combined with two different DM algorithms for optimization: the one available in the commercial TPS RayStation (RSM) and a simpler isodose-based mimicking (IBM). Using 55 test patients (five test patients for each model), we evaluated the quality and robustness of the plans generated by the four proposed KBP pipelines (PP-RSM, PP-IBM, NPP-RSM, NPP-IBM). After robust evaluation, dose-volume histogram (DVH) metrics in nominal and worst-case scenarios were compared to those of the manually generated plans. Nominal doses from the four KBP pipelines showed promising results achieving comparable target coverage and improved dose to organs at risk (OARs) compared to the manual plans. However, too optimistic post-processing applied to the dose prediction (i.e. important decrease of the dose to the organs) compromised the robustness of the plans. Even though RSM seemed to partially compensate for the lack of robustness in the PP plans, still 65% of the patients did not achieve the expected robustness levels. NPP-RSM plans seemed to achieve the best trade-off between robustness and OAR sparing. PP and DM strategies are crucial steps to generate acceptable robust and deliverable IMPT plans from ML-predicted doses. Before the clinical implementation of any KBP pipeline, the PP and DM parameters predefined by the commercial TPS need to be modified accordingly with a comprehensive feedback loop in which the robustness of the final dose calculations is evaluated. With the right choice of PP and DM parameters, KBP strategies have the potential to generate IMPT plans within clinically acceptable levels comparable to plans manually generated by dosimetrists.

Borderias-Villarroel E ，Huet Dastarac M ，Barragán-Montero AM ，Helander R ，Holmstrom M ，Geets X ，Sterpin E ... -  《-》

Automated Robust Planning for IMPT in Oropharyngeal Cancer Patients Using Machine Learning.

The aim of this study was to evaluate an automated treatment planning method for robustly optimized intensity modulated proton therapy (IMPT) plans for oropharyngeal carcinoma patients and to compare the results with manually optimized robust IMPT plans. An atlas regression forest-based machine learning (ML) model for dose prediction was trained on CT scans, contours, and dose distributions of robust IMPT plans of 88 oropharyngeal cancer (OPC) patients. The ML model was combined with a robust voxel and dose volume histogram-based dose mimicking optimization algorithm for 21 perturbed scenarios to generate a machine-deliverable plan from the predicted dose distribution. Machine learning optimization (MLO) configuration was performed using a cross-validation approach with 3 × 8 tuning patients and comprised of adjustments to the mimicking optimization, to generate higher-quality MLO plans. Independent testing of the MLO algorithm was performed with another 25 patients. Plan quality of clinical and MLO plans was evaluated by the clinical target volume (D98% voxel-wise minimum dose >94%), organ at risk (OAR) doses, and the normal tissue complication probability (NTCP) (sum (Σ) of grade-2 and grade-3 dysphagia and xerostomia). Adequate robust target coverage was achieved in 24/25 clinical plans and in 23/25 MLO plans in the primary clinical target volume (CTV). In the elective CTV, 22/25 clinical plans and 24/25 MLO plans passed the robust target coverage evaluation threshold. The MLO average Σgrade 2 and Σgrade 3 NTCPs were comparable to the clinical plans (Σgrade 2 NTCPs: clinical 47.5% vs MLO 48.4%, Σgrade 3 NTCPs: clinical 11.9% vs MLO 12.3%). Significant increases in OAR average doses in MLO plans were found in the pharynx constrictor muscles (163 cGy, P = .002) and cervical esophagus (265 cGy, P = .002). The MLO plans were created within 45 minutes. This study showed that automated MLO planning can generate robustly optimized MLO plans with quality comparable to clinical plans in OPC patients.

van Bruggen IG ，Huiskes M ，de Vette SPM ，Holmström M ，Langendijk JA ，Both S ，Kierkels RGJ ，Korevaar EW ... -  《-》

Superiority in Robustness of Multifield Optimization Over Single-Field Optimization for Pencil-Beam Proton Therapy for Oropharynx Carcinoma: An Enhanced Robustness Analysis.

To compare the difference in robustness of single-field optimized (SFO) and robust multifield optimized (rMFO) proton plans for oropharynx carcinoma patients by an improved robustness analysis. We generated rMFO proton plans for 11 patients with oropharynx carcinoma treated with SFO intensity modulated proton therapy with simultaneous integrated boost prescription. Doses from both planning approaches were compared for the initial plans and the worst cases from 20 optimization scenarios of setup errors and range uncertainties. Expected average dose distributions per range uncertainty were obtained by weighting the contributions from the respective scenarios with their expected setup error probability, and the spread of dose parameters for different range uncertainties were quantified. Using boundary dose distributions created from 56 combined setup error and range uncertainty scenarios and considering the vanishing influence of setup errors after 30 fractions, we approximated realistic worst-case values for the total treatment course. Error bar metrics derived from these boundary doses are reported for the clinical target volumes (CTVs) and organs at risk (OARs). The rMFO plans showed improved CTV coverage and homogeneity while simultaneously reducing the average mean dose to the constrictor muscles, larynx, and ipsilateral middle ear by 5.6 Gy, 2.0 Gy, and 3.9 Gy, respectively. We observed slightly larger differences during robustness evaluation, as well as a significantly higher average brainstem maximum and ipsilateral parotid mean dose for SFO plans. For rMFO plans, the range uncertainty-related spread in OAR dose parameters and many error bar metrics were found to be superior. The SFO plans showed a lower global maximum dose for single-scenario worst cases and a slightly lower mean oral cavity dose throughout. An enhanced robustness analysis has been proposed and implemented into clinical systems. The benefit of better CTV coverage and OAR dose sparing in oropharynx carcinoma patients by rMFO compared with SFO proton plans is preserved in a robustness analysis with consideration of setup error and range uncertainty.

Stützer K ，Lin A ，Kirk M ，Lin L ... -  《-》

Technical Note: Treatment planning system (TPS) approximations matter - comparing intensity-modulated proton therapy (IMPT) plan quality and robustness between a commercial and an in-house developed TPS for nonsmall cell lung cancer (NSCLC).

Approximate dose calculation methods were used in the nominal dose distribution and the perturbed dose distributions due to uncertainties in a commercial treatment planning system (CTPS) for robust optimization in intensity-modulated proton therapy (IMPT). We aimed to investigate whether the approximations influence plan quality, robustness, and interplay effect of the resulting IMPT plans for the treatment of locally advanced lung cancer patients. Ten consecutively treated locally advanced nonsmall cell lung cancer (NSCLC) patients were selected. Two IMPT plans were created for each patient using our in-house developed TPS, named "Solo," and also the CTPS, EclipseTM (Varian Medical Systems, Palo Alto, CA, USA), respectively. The plans were designed to deliver prescription doses to internal target volumes (ITV) drawn by a physician on averaged four-dimensional computed tomography (4D-CT). Solo plans were imported back to CTPS, and recalculated in CTPS for fair comparison. Both plans were further verified for each patient by recalculating doses in the inhalation and exhalation phases to ensure that all plans met clinical requirements. Plan robustness was quantified on all phases using dose-volume-histograms (DVH) indices in the worst-case scenario. The interplay effect was evaluated for every plan using an in-house developed software, which randomized starting phases of each field per fraction and accumulated dose in the exhalation phase based on the patient's breathing motion pattern and the proton spot delivery in a time-dependent fashion. DVH indices were compared using Wilcoxon rank-sum test. Compared to the plans generated using CTPS on the averaged CT, Solo plans had significantly better target dose coverage and homogeneity (normalized by the prescription dose) in the worst-case scenario [ITV D95% : 98.04% vs 96.28%, Solo vs CTPS, P = 0.020; ITV D5% -D95% : 7.20% vs 9.03%, P = 0.049] while all DVH indices were comparable in the nominal scenario. On the inhalation phase, Solo plans had better target dose coverage and cord Dmax in the nominal scenario [ITV D95% : 99.36% vs 98.45%, Solo vs CTPS, P = 0.014; cord Dmax : 20.07 vs 23.71 Gy(RBE), P = 0.027] with better target coverage and cord Dmax in the worst-case scenario [ITV D95% : 97.89% vs 96.47%, Solo vs CTPS, P = 0.037; cord Dmax : 24.57 vs 28.14 Gy(RBE), P = 0.037]. On the exhalation phase, similar phenomena were observed in the nominal scenario [ITV D95% : 99.63% vs 98.87%, Solo vs CTPS, P = 0.037; cord Dmax : 19.67 vs 23.66 Gy(RBE), P = 0.039] and in the worst-case scenario [ITV D95% : 98.20% vs 96.74%, Solo vs CTPS, P = 0.027; cord Dmax : 23.47 vs 27.93 Gy(RBE), P = 0.027]. In terms of interplay effect, plans generated by Solo had significantly better target dose coverage and homogeneity, less hot spots, and lower esophageal Dmean , and cord Dmax [ITV D95% : 101.81% vs 98.68%, Solo vs CTPS, P = 0.002; ITV D5% -D95% : 2.94% vs 7.51%, P = 0.002; cord Dmax : 18.87 vs 22.29 Gy(RBE), P = 0.014]. Solo-generated IMPT plans provide improved cord sparing, better target robustness in all considered phases, and reduced interplay effect compared with CTPS. Consequently, the approximation methods currently used in commercial TPS programs may have space for improvement in generating optimal IMPT plans for patient cases with locally advanced lung cancer.

Liu C ，Yu NY ，Shan J ，Bhangoo RS ，Daniels TB ，Chiang JS ，Ding X ，Lara P ，Patrick CL ，Archuleta JP ，DeWees T ，Hu Y ，Schild SE ，Bues M ，Sio TT ，Liu W ... -  《-》

A novel and individualized robust optimization method using normalized dose interval volume constraints (NDIVC) for intensity-modulated proton radiotherapy.

Intensity-modulated proton therapy (IMPT) is known to be sensitive to patient setup and range uncertainty issues. Multiple robust optimization methods have been developed to mitigate the impact of these uncertainties. Here, we propose a new robust optimization method, which provides an alternative way of robust optimization in IMPT, and is clinically practical, which will enable users to control the balance between nominal plan quality and plan robustness in a user-defined fashion. We calculated nine individual dose distributions which corresponded to one nominal and eight extreme scenarios caused by patient setup and proton beam's range uncertainties. For each voxel, the normalized dose interval (NDI) is defined as the full dose range variation divided by the maximum dose in all uncertainty scenarios (NDI = [max - min dose]/max dose), which was then used to calculate the normalized dose interval volume histogram (NDIVH) curves. The areas under the NDIVH curves were used to quantify plan robustness. A normalized dose interval volume constraint (NDIVC) applied to the target was incorporated to specify the desired robustness which was user-defined. Users could then explore the trade-off between nominal plan quality and plan robustness by adjusting the position of the NDIVCs on the NDIVH curves freely. We benchmarked our method using one lung, five head and neck (H&N), and three prostate cases by comparing our results to those derived using the voxel-wise worst-case robust optimization. Using the benchmark cases, our new method achieved quality IMPT plans comparable to those derived from the voxel-wise worst-case robust optimization for both nominal plan quality and plan robustness in general; even more conformal and more homogeneous target dose distributions in some cases, if proper NDIVCs were applied. The AUC under NDIVH, as a precise quantitative index of plan robustness, was consistent with DVH bandwidths. Additionally, we demonstrated the feasibility of adjusting the position of NDIVCs in the NDIVH curves which allowed users to explore the trade-off between nominal plan quality and plan robustness. The NDIVH-based robust optimization method provided a novel and individualized way of robust optimization in IMPT, and enables users to adjust the balance between nominal plan quality and plan robustness in a user-defined fashion. This method is applicable for continued improvement and developing the next generation of IMPT planning algorithms in the future.

Shan J ，Sio TT ，Liu C ，Schild SE ，Bues M ，Liu W ... -  《-》