Pareto-optimal plans as ground truth for validation of a commercial system for knowledge-based DVH-prediction

Dosimetry and efficiency comparison of knowledge-based and manual planning using volumetric modulated arc therapy for craniospinal irradiation

BACKGROUND

Craniospinal irradiation (CSI) poses a challenge to treatment planning due to the large target, field junction, and multiple organs at risk (OARs) involved. The aim of this study was to evaluate the performance of knowledge-based planning (KBP) in CSI by comparing original manual plans (MP), KBP RapidPlan initial plans (RPI), and KBP RapidPlan final plans (RPF), which received further re-optimization to meet the dose constraints.

PATIENTS AND METHODS

Dose distributions in the target were evaluated in terms of coverage, mean dose, conformity index (CI), and homogeneity index (HI). The dosimetric results of OARs, planning time, and monitor unit (MU) were evaluated.

RESULTS

All MP and RPF plans met the plan goals, and 89.36% of RPI plans met the plan goals. The Wilcoxon tests showed comparable target coverage, CI, and HI for the MP and RPF groups; however, worst plan quality was demonstrated in the RPI plans than in MP and RPF. For the OARs, RPF and RPI groups had better dosimetric results than the MP group (P < 0.05 for optic nerves, eyes, parotid glands, and heart). The planning time was significantly reduced by the KBP from an average of 677.80 min in MP to 227.66 min (P < 0.05) and 307.76 min (P < 0.05) in RPI, and RPF, respectively. MU was not significantly different between these three groups.

CONCLUSIONS

The KBP can significantly reduce planning time in CSI. Manual re-optimization after the initial KBP is recommended to enhance the plan quality.

Determining the library size for the optimal output plan in the RapidPlan knowledge-based planning system using multicriteria optimization

OBJECTIVES

The aim of this study was to determine the number of trade-off explored (TO) library plans required for building a RapidPlan (RP) library that would generate the optimal clinical treatment plan.

METHODS

We developed 2 RP models, 1 each for the 2 clinical sites, head and neck (HN) and cervix. The models were created using 100 plans and were validated using 70 plans (VP) for each site respectively. Each of the 2 libraries comprising 100 TO plans was divided into 5 different subsets of library plans comprising 20, 40, 60, 80, and 100 plans, leading to 5 different RP models for each site. For every validation patient, a TO plan (TO_VP) was created. For every patient, 5 RP plans were automatically generated using RP models. The dosimetric parameters of the 6 plans (TO_VP + 5 RP plans) were compared using Pearson correlation and Greenhouse-Geisser analysis.

RESULTS

Planning target volume (PTV) dose volume parameters PTVD95% in 6 competing plans varied between 97.6 ± 0.7% and 98.1 ± 0.6% in HN cases and 98.8 ± 0.3% and 99.0 ± 0.4% in cervix cases. Overall, for both sites, the mean variations in organ at risk (OAR) doses or volumes were within 50 cGy, 0.5%, and 0.2 cc between library plans, and if TO_VP was included the variations deteriorated to 180 cGy, 0.4%, and 15 cc. All OARs in both sites, except D0.1 ccspine, showed a statistically insignificant variation between all plans.

CONCLUSIONS

Dosimetric variation among various output plans generated from 5 RP libraries is minimal and clinically insignificant. The optimal output plan can be derived from the least-weighted library consisting of 20 plans.

ADVANCES IN KNOWLEDGE

This article shows that, when the constituent plans are subjected to trade-off exploration, the number of constituent plans for a knowledge-based planning module is not relevant in terms of its dosimetric output.

Knowledge-based versus deep learning based treatment planning for breast radiotherapy

Background and Purpose

To improve radiotherapy (RT) planning efficiency and plan quality, knowledge-based planning (KBP) and deep learning (DL) solutions have been developed. We aimed to make a direct comparison of these models for breast cancer planning using the same training, validation, and testing sets.

Materials and Methods

Two KBP models were trained and validated with 90 RT plans for left-sided breast cancer with 15 fractions of 2.6 Gy. The versions either used the full dataset (non-clean model) or a cleaned dataset (clean model), thus eliminating geometric and dosimetric outliers. Results were compared with a DL U-net model (previously trained and validated with the same 90 RT plans) and manually produced RT plans, for the same independent dataset of 15 patients. Clinically relevant dose volume histogram parameters were evaluated according to established consensus criteria.

Results

Both KBP models underestimated the mean heart and lung dose equally 0.4 Gy (0.3-1.1 Gy) and 1.4 Gy (1.1-2.8 Gy) compared to the clinical plans 0.8 Gy (0.5-1.8 Gy) and 1.7 Gy (1.3-3.2 Gy) while in the final calculations the mean lung dose was higher 1.9-2.0 Gy (1.5-3.5 Gy) for both KPB models. The U-Net model resulted in a mean planning target volume dose of 40.7 Gy (40.4-41.3 Gy), slightly higher than the clinical plans 40.5 Gy (40.1-41.0 Gy).

Conclusions

Only small differences were observed between the estimated and final dose calculation and the clinical results for both KPB models and the DL model. With a good set of breast plans, the data cleaning module is not needed and both KPB and DL models lead to clinically acceptable results.

Knowledge-based plan optimization for prostate SBRT delivered with CyberKnife according to RTOG0938 protocol

PURPOSE

To extend the knowledge-based (KB) automatic planning approach to CyberKnife in the case of Stereotactic Body Radiation Therapy (SBRT) for prostate cancer.

METHODS

Seventy-two clinical plans of patients treated according to the RTOG0938 protocol (36.25 Gy/5fr) with CyberKnife were exported from the CyberKnife system to Eclipse to train a KB-model using the Rapid Plan tool. The KB approach provided dose-volume objectives for specific OARs only and not PTV. Bladder, rectum and femoral heads were considered in the model. The KB-model was successfully trained on 51 plans and then validated on 20 new patients. A KB-based template was tuned in the Precision system for both sequential optimization (SO) and VOLO optimization algorithms. Plans of the validation group were re-optimized (KB-TP) using both algorithms without any operator intervention and compared against the original plans (TP) in terms of OARs/PTV dose-volume parameters. Paired Wilcoxon signed-rank tests were performed to assess statistically significant differences (p < 0.05).

RESULTS

Regarding SO, automatic KB-TP plans were generally better than or equivalent to TP plans. PTVs V95% was slightly worse while OARs sparing for KB-TP was significantly improved. Regarding VOLO optimization, the PTVs coverage was significantly better for KB-TP while there was a limited worsening in the rectum. A significant improvement was observed in the bladder in the range of low-intermediate doses.

CONCLUSIONS

An extension of the KB optimization approach to the CyberKnife system has been successfully developed and validated in the case of SBRT prostate cancer.

Attention-aware 3D U-Net convolutional neural network for knowledge-based planning 3D dose distribution prediction of head-and-neck cancer

PURPOSE

Deep learning-based knowledge-based planning (KBP) methods have been introduced for radiotherapy dose distribution prediction to reduce the planning time and maintain consistent high-quality plans. This paper presents a novel KBP model using an attention-gating mechanism and a three-dimensional (3D) U-Net for intensity-modulated radiation therapy (IMRT) 3D dose distribution prediction in head-and-neck cancer.

METHODS

A total of 340 head-and-neck cancer plans, representing the OpenKBP-2020 AAPM Grand Challenge data set, were used in this study. All patients were treated with the IMRT technique and a dose prescription of 70 Gy. The data set was randomly divided into 64%/16%/20% as training/validation/testing cohorts. An attention-gated 3D U-Net architecture model was developed to predict full 3D dose distribution. The developed model was trained using the mean-squared error loss function, Adam optimization algorithm, a learning rate of 0.001, 120 epochs, and batch size of 4. In addition, a baseline U-Net model was also similarly trained for comparison. The model performance was evaluated on the testing data set by comparing the generated dose distributions against the ground-truth dose distributions using dose statistics and clinical dosimetric indices. Its performance was also compared to the baseline model and the reported results of other deep learning-based dose prediction models.

RESULTS

The proposed attention-gated 3D U-Net model showed high capability in accurately predicting 3D dose distributions that closely replicated the ground-truth dose distributions of 68 plans in the test set. The average value of the mean absolute dose error was 2.972 ± 1.220 Gy (vs. 2.920 ± 1.476 Gy for a baseline U-Net) in the brainstem, 4.243 ± 1.791 Gy (vs. 4.530 ± 2.295 Gy for a baseline U-Net) in the left parotid, 4.622 ± 1.975 Gy (vs. 4.223 ± 1.816 Gy for a baseline U-Net) in the right parotid, 3.346 ± 1.198 Gy (vs. 2.958 ± 0.888 Gy for a baseline U-Net) in the spinal cord, 6.582 ± 3.748 Gy (vs. 5.114 ± 2.098 Gy for a baseline U-Net) in the esophagus, 4.756 ± 1.560 Gy (vs. 4.992 ± 2.030 Gy for a baseline U-Net) in the mandible, 4.501 ± 1.784 Gy (vs. 4.925 ± 2.347 Gy for a baseline U-Net) in the larynx, 2.494 ± 0.953 Gy (vs. 2.648 ± 1.247 Gy for a baseline U-Net) in the PTV_70, and 2.432 ± 2.272 Gy (vs. 2.811 ± 2.896 Gy for a baseline U-Net) in the body contour. The average difference in predicting the D₉₉ value for the targets (PTV_70, PTV_63, and PTV_56) was 2.50 ± 1.77 Gy. For the organs at risk, the average difference in predicting the D m a x ${D_{max}}$ (brainstem, spinal cord, and mandible) and D m e a n ${D_{mean}}$ (left parotid, right parotid, esophagus, and larynx) values was 1.43 ± 1.01 and 2.44 ± 1.73 Gy, respectively. The average value of the homogeneity index was 7.99 ± 1.45 for the predicted plans versus 5.74 ± 2.95 for the ground-truth plans, whereas the average value of the conformity index was 0.63 ± 0.17 for the predicted plans versus 0.89 ± 0.19 for the ground-truth plans. The proposed model needs less than 5 s to predict a full 3D dose distribution of 64 × 64 × 64 voxels for a new patient that is sufficient for real-time applications.

CONCLUSIONS

The attention-gated 3D U-Net model demonstrated a capability in predicting accurate 3D dose distributions for head-and-neck IMRT plans with consistent quality. The prediction performance of the proposed model was overall superior to a baseline standard U-Net model, and it was also competitive to the performance of the best state-of-the-art dose prediction method reported in the literature. The proposed model could be used to obtain dose distributions for decision-making before planning, quality assurance of planning, and guiding-automated planning for improved plan consistency, quality, and planning efficiency.

Probabilistic Pareto plan generation for semiautomated multicriteria radiation therapy treatment planning

Objective.We propose a semiautomatic pipeline for radiation therapy treatment planning, combining ideas from machine learning-automated planning and multicriteria optimization (MCO).Approach.Using knowledge extracted from historically delivered plans, prediction models for spatial dose and dose statistics are trained and furthermore systematically modified to simulate changes in tradeoff priorities, creating a set of differently biased predictions. Based on the predictions, an MCO problem is subsequently constructed using previously developed dose mimicking functions, designed in such a way that its Pareto surface spans the range of clinically acceptable yet realistically achievable plans as exactly as possible. The result is an algorithm outputting a set of Pareto optimal plans, either fluence-based or machine parameter-based, which the user can navigate between in real time to make adjustments before a final deliverable plan is created.Main results.Numerical experiments performed on a dataset of prostate cancer patients show that one may often navigate to a better plan than one produced by a single-plan-output algorithm.Significance.We demonstrate the potential of merging MCO and a data-driven workflow to automate labor-intensive parts of the treatment planning process while maintaining a certain extent of manual control for the user.

Implementation of automatic plan optimization in Italy: Status and perspectives

PURPOSE

To investigate and report on the diffusion and clinical use of automated radiotherapy planning systems in Italy and to assess the perspectives of the community of Italian medical physicists involved in radiotherapy on the use of these tools.

MATERIALS AND METHODS

A survey of medical physicists (one per Institute) of 175 radiotherapy centers in Italy was conducted between February 21st and April 1st, 2021. The information collected included the institute's characteristics, plan activity, availability/use of automatic tools and related issues regarding satisfaction, criticisms, expectations, and perceived professional modifications. Responses were analysed, including the impact of a few variables such as the institute type and experience.

RESULTS

125 of the centers (71%) answered the survey, with regional variability (range: 47%-100%); among these, 49% have a TPS with some automatic option. Clinical use of automatic planning is present in 33% of the centers, with 13% applying it in >50% of their plans. Among the 125 responding centres the most used systems are Pinnacle (16%), Raystation (9%) and Eclipse (4%). The majority of participants consider the use of automated techniques to be beneficial, while only 1% do not see any advantage; 83% of respondents see the possibility of enriching their professional role as a potential benefit, while 3% see potential threats.

CONCLUSIONS

Our survey shows that 49% of the responding centres have an automatic planning solution although clinically used in only 33% of the cases. Most physicists consider the use of automated techniques to be beneficial and show a prevalently positive attitude.

Variations in Head and Neck Treatment Plan Quality Assessment Among Radiation Oncologists and Medical Physicists in a Single Radiotherapy Department

Background

Agreement between planners and treating radiation oncologists (ROs) on plan quality criteria is essential for consistent planning. Differences between ROs and planning medical physicists (MPs) in perceived quality of head and neck cancer plans were assessed.

Materials and Methods

Five ROs and four MPs scored 65 plans for in total 15 patients. For each patient, the clinical (CLIN) plan and two or four alternative plans, generated with automated multi-criteria optimization (MCO), were included. There was always one MCO plan aiming at maximally adhering to clinical plan requirements, while the other MCO plans had a lower aimed quality. Scores were given as follows: 1-7 and 1-2, not acceptable; 3-5, acceptable if further planning would not resolve perceived weaknesses; and 6-7, straightway acceptable. One MP and one RO repeated plan scoring for intra-observer variation assessment.

Results

For the 36 unique observer pairs, the median percentage of plans for which the two observers agreed on a plan score (100% = 65 plans) was 27.7% [6.2, 40.0]. In the repeat scoring, agreements between first and second scoring were 52.3% and 40.0%, respectively. With a binary division between unacceptable (scores 1 and 2) and acceptable (3-7) plans, the median inter-observer agreement percentage was 78.5% [63.1, 86.2], while intra-observer agreements were 96.9% and 86.2%. There were no differences in observed agreements between RO-RO, MP-MP, and RO-MP pairs. Agreements for the highest-quality, automatically generated MCO plans were higher than for the CLIN plans.

Conclusions

Inter-observer differences in plan quality scores were substantial and could result in inconsistencies in generated treatment plans. Agreements among ROs were not better than between ROs and MPs, despite large differences in training and clinical role. High-quality automatically generated plans showed the best score agreements.

Replacing Manual Planning of Whole Breast Irradiation With Knowledge-Based Automatic Optimization by Virtual Tangential-Fields Arc Therapy

Purpose

To implement Knowledge Based (KB) automatic planning for right and left-sided whole breast treatment through a new volumetric technique (ViTAT, Virtual Tangential-fields Arc Therapy) mimicking conventional tangential fields (TF) irradiation.

Materials and Method

A total of 193 clinical plans delivering TF with wedged or field-in-field beams were selected to train two KB-models for right(R) and left(L) sided breast cancer patients using the RapidPlan (RP) tool implemented in the Varian Eclipse system. Then, a template for ViTAT optimization, incorporating individual KB-optimized constraints, was interactively fine-tuned. ViTAT plans consisted of four arcs (6 MV) with start/stop angles consistent with the TF geometry variability within our population; the delivery was completely blocked along the arcs, apart from the first and last 20° of rotation for each arc. Optimized fine-tuned KB templates for automatic plan optimization were generated. Validation tests were performed on 60 new patients equally divided in R and L breast treatment: KB automatic ViTAT-plans (KB-ViTAT) were compared against the original TF plans in terms of OARs/PTVs dose-volume parameters. Wilcoxon-tests were used to assess the statistically significant differences.

Results

KB models were successfully generated for both L and R sides. Overall, 1(3%) and 7(23%) out of 30 automatic KB-ViTAT plans were unacceptable compared to TF for R and L side, respectively. After the manual refinement of the start/stop angles, KB-ViTAT plans well fitted TF-performances for these patients as well. PTV coverage was comparable, while PTV D_1% was improved with KB-ViTAT by R:0.4/L:0.2 Gy (p < 0.05); ipsilateral OARs D_mean were similar with a slight (i.e., few % volume) improvement/worsening in the 15–35 Gy/2–15 Gy range, respectively. KB-ViTAT better spared contralateral OARs: D_mean of contralateral OARs was 0.1 Gy lower (p < 0.05); integral dose was R:5%/L:8% lower (p < 0.05) than TF. The overall time for the automatic plan optimization and final dose calculation was 12 ± 2 minutes.

Conclusions

Fully automatic KB-optimization of ViTAT can efficiently replace manually optimized TF planning for whole breast irradiation. This approach was clinically implemented in our institute and may be suggested as a large-scale strategy for efficiently replacing manual planning with large sparing of time, elimination of inter-planner variability and of, seldomly occurring, sub-optimal manual plans.

Use of knowledge based DVH predictions to enhance automated re-planning strategies in head and neck adaptive radiotherapy

This study aimed to investigate if a commercial, knowledge-based tool for radiotherapy planning could be used to estimate the amount of sparing in organs at risk (OARs) in the re-planning strategy for adaptive radiotherapy (ART). Eighty head and neck (HN) VMAT Pareto plans from our institute's database were used to train a knowledge-based planning (KBP) model. An evaluation set of another 20 HN patients was randomly selected. For each patient in the evaluation set, the planning computed tomography (CT) and 2 sets of on-board cone-beam CT, corresponding to the middle and second half of the radiotherapy treatment course, were extracted. The original plan was re-calculated on a daily deformed CT (delivered dose-volume histogram (DVH)) and compared with the KBP DVH predictions and with the final KBP DVH after optimisation of the plan, which was performed on the same image sets. To evaluate the feasibility of this method, the range of KBP DVH uncertainties was compared with the gains obtained from re-planning. DVH differences and receiver operating characteristic (ROC) curve analysis were used for this purpose. On average, final KBP uncertainties were smaller than the gain in re-planning. Statistical tests confirmed significant differences between the two groups. ROC analysis showed KBP performance in terms of area under the curve values higher than 0.7, which confirmed a good accuracy in predicted values. Overall, for 48% of cases, KBP predicted a desirable outcome from re-planning, and the final dose confirmed an effective gain in 47% of cases. We have established a systematic workflow to identify effective OAR sparing in re-planning based on KBP predictions that can be implemented in an on-line, ART process.

Assessment of a conventional volumetric-modulated arc therapy knowledge-based planning model applied to the new Halcyon© O-ring linac in locoregional breast cancer radiotherapy

INTRODUCTION

The aim of this study was to evaluate the performance of a knowledge-based planning (KBP) model for breast cancer trained on plans performed on a conventional linac with 6 MV FF (flattening filter) beams and volumetric-modulated arc therapy (VMAT) for plans performed on the new jawless Halcyon© system with 6 MV FFF (flattening filter-free) beams.

MATERIALS AND METHODS

Based on the RapidPlan© (RP) KBP optimization engine, a DVH Estimation Model was first trained using 56 VMAT left-sided breast cancer treatment plans performed on a conventional linac, and validated on another 20 similar cases (without manual intervention). To determine the capacity of the model for Halcyon©, an additional cohort of 20 left-sided breast cancer plans was generated with RP and analyzed for both TrueBeam© and Halcyon© machines. Plan qualities between manual vs RP (followed by manual intervention) Halcyon© plans set were compared qualitatively by blinded review by radiation oncologists for 10 new independent plans.

RESULTS

Halcyon© plans generated with the VMAT model trained with conventional linac plans showed comparable target dose distribution compared to TrueBeam© plans. Organ sparingwas comparable between the 2 devices with a slight decrease in heart dose for Halcyon© plans. Nine out of ten automatically generated Halcyon© plans were preferentially chosen by the radiation oncologists over the manually generated Halcyon© plans.

CONCLUSION

A VMAT KBP model driven by plans performed on a conventional linac with 6 MV FF beams provides high quality plans performed with 6 MV FFF beams on the new Halcyon© linac.

Artificial Intelligence and the Medical Physicist: Welcome to the Machine

Artificial intelligence (AI) is a branch of computer science dedicated to giving machines or computers the ability to perform human-like cognitive functions, such as learning, problem-solving, and decision making. Since it is showing superior performance than well-trained human beings in many areas, such as image classification, object detection, speech recognition, and decision-making, AI is expected to change profoundly every area of science, including healthcare and the clinical application of physics to healthcare, referred to as medical physics. As a result, the Italian Association of Medical Physics (AIFM) has created the “AI for Medical Physics” (AI4MP) group with the aims of coordinating the efforts, facilitating the communication, and sharing of the knowledge on AI of the medical physicists (MPs) in Italy. The purpose of this review is to summarize the main applications of AI in medical physics, describe the skills of the MPs in research and clinical applications of AI, and define the major challenges of AI in healthcare.

Evaluation of a generalized knowledge-based planning performance for VMAT irradiation of breast and locoregional lymph nodes-Internal mammary and/or supraclavicular regions

PURPOSE

To evaluate the performance of eleven Knowledge-Based (KB) models for planning optimization (RapidPlantm (RP), Varian) of Volumetric Modulated Arc Therapy (VMAT) applied to whole breast comprehensive of nodal stations, internal mammary and/or supraclavicular regions.

METHODS AND MATERIALS

Six RP models have been generated and trained based on 120 VMAT plans data set with different criteria. Two extra-structures were delineated: a PTV for the optimization and a ring structure. Five more models, twins of the previous models, have been created without the need of these structures.

RESULTS

All models were successfully validated on an independent cohort of 40 patients, 30 from the same institute that provided the training patients and 10 from an additional institute, with the resulting plans being of equal or better quality compared with the clinical plans. The internal validation shows that the models reduce the heart maximum dose of about 2 Gy, the mean dose of about 1 Gy and the V20Gy of 1.5 Gy on average. Model R and L together with model B without optimization structures ensured the best outcomes in the 20% of the values compared to other models. The external validation observed an average improvement of at least 16% for the V5Gy of lungs in RP plans. The mean heart dose and for the V20Gy for lung IPSI were almost halved. The models reduce the maximum dose for the spinal canal of more than 2 Gy on average.

CONCLUSIONS

All KB models allow a homogeneous plan quality and some dosimetric gains, as we saw in both internal and external validation. Sub-KB models, developed by splitting right and left breast cases or including only whole breast with locoregional lymph nodes, have shown good performances, comparable but slightly worse than the general model. Finally, models generated without the optimization structures, performed better than the original ones.

Clinical Implementation of Knowledge-Based Automatic Plan Optimization for Helical Tomotherapy

PURPOSE

To implement knowledge-based (KB) automatic planning for helical TomoTherapy (HTT). The focus of the first clinical implementation was the case of high-risk prostate cancer, including pelvic node irradiation.

METHODS AND MATERIALS

One hundred two HTT clinical plans were selected to train a KB model using the RapidPlan tool incorporated in the Eclipse system (v13.6, Varian Inc). The individually optimized KB-based templates were converted into HTT-like templates and sent automatically to the HTT treatment planning system through scripting. The full dose calculation was set after 300 iterations without any additional planner intervention. Internal (20 patients in the training cohort) and external (28 new patients) validation were performed to assess the performance of the model: Automatic HTT plans (KB-TP) were compared against the original plans (TP) in terms of organs at risk and planning target volume (PTV) dose-volume parameters and by blinded clinical evaluation of 3 expert clinicians.

RESULTS

KB-TP plans were generally better than or equivalent to TP plans in both validation cohorts. A significant improvement in PTVs and rectum-PTV overlap dosimetry parameters were observed for both sets. Organ-at-risk sparing for KB-TP was slightly improved, which was more evident in the external validation group and for bladder and bowel. Clinical evaluation reported KB-TP to be better in 60% of cases and worse in 10% compared with TP (P < .05).

CONCLUSIONS

The fully KB-based automatic planning workflow was successfully implemented for HTT planning optimization in the case of high-risk patients with prostate cancer.

Predicting patient specific Pareto fronts from patient anatomy only

PURPOSE

To demonstrate the feasibility of predicting the patient-specific treatment planning Pareto front (PF) for prostate cancer patients based only on delineations of PTV, rectum and body.

MATERIAL/METHODS

Our methodology consists of four steps. First, using Erasmus-iCycle, the Pareto fronts of 112 prostate cancer patients were constructed by generating per patient 42 Pareto optimal treatment plans with different priorities. Dose parameters associated to homogeneity, conformity and dose to rectum were extracted. Second, a 3D convex function representing the PF spanned by the 42 plans was fitted for each patient using three patient-specific parameters. Third, ten features were extracted from the, aforementioned, structures to train a linear-regressor prediction algorithm to predict these three patient-specific parameters. Fourth, the quality of the predictions was assessed by calculating the average and maximum distances of the predicted PF to the 42 plans for patients in the validation cohort.

RESULTS

The prediction model was able to predict the clinically relevant PF within 2 Gy for 90% of the patients with a median average distance of 0.6 Gy.

CONCLUSIONS

We demonstrate the feasibility of fast, accurate predictions of the patient-specific PF for prostate cancer patients based only on delineations of PTV, rectum and body.

Knowledge-based automatic optimization of adaptive early-regression-guided VMAT for rectal cancer

PURPOSE

To implement a knowledge-based (KB) optimization strategy to our adaptive (ART) early-regression guided boosting technique in neo-adjuvant radio-chemotherapy for rectal cancer.

MATERIAL AND METHODS

The protocol consists of a first phase delivering 27.6 Gy to tumor/lymph-nodes (2.3 Gy/fr-PTV₁), followed by the ART phase concomitantly delivering 18.6 Gy (3.1 Gy/fr) and 13.8 Gy (2.3 Gy/fr) to the residual tumor (PTV_ART) and to PTV₁ respectively. PTV_ART is obtained by expanding the residual GTV, as visible on MRI at fraction 9. Forty plans were used to generate a KB-model for the first phase using the RapidPlan tool. Instead of building a new model, a robust strategy scaling the KB-model to the ART phase was applied. Both internal and external validation were performed for both phases: all automatic plans (RP) were compared in terms of OARs/PTVs parameters against the original plans (RA).

RESULTS

The resulting automatic plans were generally better than or equivalent to clinical plans. Of note, V_30Gy and V_40Gy were significantly improved in RP plans for bladder and bowel; gEUD analysis showed improvement for KB-modality for all OARs, up to 3 Gy for the bowel.

CONCLUSIONS

The KB-model generated for the first phase was robust and it was also efficiently adapted to the ART phase. The performance of automatically generated plans were slightly better than the corresponding manual plans for both phases.

RapidPlan knowledge based planning: iterative learning process and model ability to steer planning strategies

Purpose

To determine if the performance of a knowledge based RapidPlan (RP) planning model could be improved with an iterative learning process, i.e. if plans generated by an RP model could be used as new input to re-train the model and achieve better performance.

Methods

Clinical VMAT plans from 83 patients presenting with head and neck cancer were selected to train an RP model, CL-1. With this model, new plans on the same patients were generated, and subsequently used as input to train a novel model, CL-2. Both models were validated on a cohort of 20 patients and dosimetric results compared. Another set of 83 plans was realised on the same patients with different planning criteria, by using a simple template with no attempt to manually improve the plan quality. Those plans were employed to train another model, TP-1. The differences between the plans generated by CL-1 and TP-1 for the validation cohort of patients were compared with respect to the differences between the original plans used to build the two models.

Results

The CL-2 model presented an improvement relative to CL-1, with higher R² values and better regression plots. The mean doses to parallel organs decreased with CL-2, while D_1% to serial organs increased (but not significantly). The different models CL-1 and TP-1 were able to yield plans according to each original strategy.

Conclusion

A refined RP model allowed the generation of plans with improved quality, mostly for parallel organs at risk and, possibly, also the intrinsic model quality.

A convolutional neural network approach for IMRT dose distribution prediction in prostate cancer patients

The purpose of the study was to compare a 3D convolutional neural network (CNN) with the conventional machine learning method for predicting intensity-modulated radiation therapy (IMRT) dose distribution using only contours in prostate cancer. In this study, which included 95 IMRT-treated prostate cancer patients with available dose distributions and contours for planning target volume (PTVs) and organs at risk (OARs), a supervised-learning approach was used for training, where the dose for a voxel set in the dataset was defined as the label. The adaptive moment estimation algorithm was employed for optimizing a 3D U-net similar network. Eighty cases were used for the training and validation set in 5-fold cross-validation, and the remaining 15 cases were used as the test set. The predicted dose distributions were compared with the clinical dose distributions, and the model performance was evaluated by comparison with RapidPlan™. Dose-volume histogram (DVH) parameters were calculated for each contour as evaluation indexes. The mean absolute errors (MAE) with one standard deviation (1SD) between the clinical and CNN-predicted doses were 1.10% ± 0.64%, 2.50% ± 1.17%, 2.04% ± 1.40%, and 2.08% ± 1.99% for D2, D98 in PTV-1 and V65 in rectum and V65 in bladder, respectively, whereas the MAEs with 1SD between the clinical and the RapidPlan™-generated doses were 1.01% ± 0.66%, 2.15% ± 1.25%, 5.34% ± 2.13% and 3.04% ± 1.79%, respectively. Our CNN model could predict dose distributions that were superior or comparable with that generated by RapidPlan™, suggesting the potential of CNN in dose distribution prediction.

Multi-institutional evaluation of knowledge-based planning performance of volumetric modulated arc therapy (VMAT) for head and neck cancer

PURPOSE

The aim of this study was to investigate whether additional manual objectives are necessary for the RapidPlan (RP) with a single optimization. We conducted multi-institutional comparisons of plan quality for head and neck cancer (HNC) using the models created at each institute.

METHODS

The ability of RP to produce acceptable plans for dose requirements was evaluated in two types of oropharynx cancers at five institutes in Japan. Volumetric modulated arc therapy plans created without (RP plan) and with additional manual objectives (M-RP plan) were compared in terms of planning target volume (PTV), brainstem, spinal cord and parotid glands in dosimetric parameters.

RESULTS

There were no major dosimetric PTV differences between RP and M-RP plans. For the brainstem and spinal cord in the RP plans, only 40% and 30% of the plans achieved the dose requirements, while the M-RP plans with upper objective added to volume 0% at all institutes achieved them for 90% of the plans. For the L-parotid gland, there was no difference in the RP and M-RP plans (both were 40%) in achieving the acceptable criteria. For the R-parotid gland, 60% and 80% of the RP and M-RP plans achieved the constraint criteria, and in terms of the achievement rate, the RP plans were relatively high.

CONCLUSIONS

M-RP plans did not require reoptimization; only an upper objective was needed for the brainstem and spinal cord, while the parotid gland dose was reduced in both RP plans with the auto generated line objectives alone.

Comprehensive Intra-Institution stepping validation of knowledge-based models for automatic plan optimization

PURPOSE

To develop and apply a stepping approach for the validation of Knowledge-based (KB) models for planning optimization: the method was applied to the case of concomitant irradiation of pelvic nodes and prostate + seminal-vesicles bed irradiation in post-prostatectomy patients.

METHODS

The clinical VMAT plans of 52 patients optimized by two reference planners were selected to generate a KB-model (RapidPlan, v.13.5 Varian). A stepping-validation approach was followed by comparing KB-generated plans (with and without planner-interaction, RP and only-RP respectively) against delivered clinical plans (RA). The validation followed three steps, gradually extending its generalization: 20 patients used to develop the model (closed-loop); 20 new patients, same planners (open-loop); 20 new patients, different planners (wide-loop). All plans were compared, in terms of relevant dose-volume parameters and generalized equivalent uniform dose (gEUD).

RESULTS

KB-plans were generally better than or equivalent to clinical plans. For RPvsRA, PTVs coverage was comparable, for OARs RP was always better. Comparing only-RPvsRA, PTVs coverage was always better; bowel\bladder V_50Gy and D_1%, bowel\bladder\rectum D_mean, femoral heads V_40Gy and penile bulb V_50Gy were significantly improved. For RPvsRA gEUD reduction >1 Gy was seen in 80% of plans for rectum, bladder and bowel; for only-RPvsRA, this was found in 50% for rectum/bladder and in 70% for bowel.

CONCLUSION

An extensive stepping validation approach of KB-model for planning optimization showed better or equal performances of automatically generated KB-plan compared to clinical plans. The interaction of a planner further improved planning performances.