Deep Learning-Based Delineation of Head and Neck Organs at Risk: Geometric and Dosimetric Evaluation.

PURPOSE: Organ-at-risk (OAR) delineation is a key step in treatment planning but can be time consuming, resource intensive, subject to variability, and dependent on anatomical knowledge. We studied deep learning (DL) for automated delineation of multiple OARs; in addition to geometric evaluation, the dosimetric impact of using DL contours for treatment planning was investigated. METHODS AND MATERIALS: The following OARs were delineated with DL developed in-house: both submandibular and parotid glands, larynx, cricopharynx, pharyngeal constrictor muscle (PCM), upper esophageal sphincter, brain stem, oral cavity, and esophagus. DL contours were benchmarked against the manual delineation (MD) clinical contours using the Sørensen-Dice similarity coefficient. Automated knowledge-based treatment plans were used. The mean dose to the manually delineated OAR structures was reported for the MD and DL plans. RESULTS: DL delineation of all OARs took <10 seconds per patient. For 7 of 11 OARs, the average Sørensen-Dice similarity coefficient was good (0.78-0.83). However, performance was lower for the esophagus (0.60), brainstem (0.64), PCM (0.68), and cricopharynx (0.73), often because of variations in MD. Although the average dose was statistically significantly higher in the DL plans for the inferior PCM (1.4 Gy) and esophagus (2.2 Gy), these average differences were not clinically significant. Dose to 28 of 209 (13.4%) and 7 of 209 (3.3%) OARs was >2 Gy higher and >2 Gy lower, respectively, in the DL plans. CONCLUSIONS: DL-based segmentation for head and neck OARs is fast; for most organs and most patients, it performs sufficiently well for treatment-planning purposes. It has the potential to increase efficiency and facilitate online adaptive radiation therapy.

Knowledge-Based Planning for Identifying High-Risk Stereotactic Ablative Radiation Therapy Treatment Plans for Lung Tumors Larger Than 5 cm.

PURPOSE: Stereotactic ablative body radiation therapy (SABR) for lung tumors ≥5 cm can be associated with more toxicity than that for smaller tumors. We investigated the relationship between dosimetry and toxicity and used a knowledge-based planning solution to retrospectively perform individualized treatment plan quality assurance (QA) with the aim of identifying where planning could have been improved. METHODS AND MATERIALS: Previous retrospective analysis of 53 patients with primary or recurrent non-small cell lung cancer ≥5 cm, treated with 5- or 8-fraction volumetric modulated arc therapy SABR between 2008 and 2014, showed 30% with grade ≥3 toxicity. During this period, several improvements were made to departmental planning protocols. RapidPlan was used to compare dosimetry of patients with or without grade ≥3 toxicity. A model comprising plans from patients without toxicity and compliant with the current planning protocol was used to provide QA for the plans from patients who had toxicity. RESULTS: Sixteen of 53 patients had grade ≥3 toxicity, including 10 with radiation pneumonitis (RP), 3 with lung hemorrhage (1 of these also had RP), and 1 with airway stenosis/atelectasis. RP was again shown to be significantly correlated with contralateral and total-lung V5 and mean lung dose. The 4 highest contralateral-lung doses belonged to patients with RP. Five of 10 clinical plans in patients with RP had a contralateral-lung mean dose up to 2.5 times higher than that of the knowledge-based plan. For 2 of 3 patients with lung hemorrhage and 1 with airway stenosis/atelectasis, the clinical plans had the highest proximal bronchial tree doses, which was also higher than in plans from the model. In 8 patients with grade ≥3 toxicity, clinical plans had dosimetry similar to that in the predictions from the model. CONCLUSIONS: A "no-toxicity" RapidPlan model identified the potential for dosimetric improvement in nearly 50% of historical treatment plans from patients with grade ≥3 toxicity after SABR for lung tumors ≥5 cm. Model-based QA may be useful for benchmarking treatment planning protocols in routine practice and in clinical studies.

Automated Knowledge-Based Intensity-Modulated Proton Planning: An International Multicenter Benchmarking Study.

Background: Radiotherapy treatment planning is increasingly automated and knowledge-based planning has been shown to match and sometimes improve upon manual clinical plans, with increased consistency and efficiency. In this study, we benchmarked a novel prototype knowledge-based intensity-modulated proton therapy (IMPT) planning solution, against three international proton centers. Methods: A model library was constructed, comprising 50 head and neck cancer (HNC) manual IMPT plans from a single center. Three external-centers each provided seven manual benchmark IMPT plans. A knowledge-based plan (KBP) using a standard beam arrangement for each patient was compared with the benchmark plan on the basis of planning target volume (PTV) coverage and homogeneity and mean organ-at-risk (OAR) dose. Results: PTV coverage and homogeneity of KBPs and benchmark plans were comparable. KBP mean OAR dose was lower in 32/54, 45/48 and 38/53 OARs from center-A, -B and -C, with 23/32, 38/45 and 23/38 being >2 Gy improvements, respectively. In isolated cases the standard beam arrangement or an OAR not being included in the model or being contoured differently, led to higher individual KBP OAR doses. Generating a KBP typically required <10 min. Conclusions: A knowledge-based IMPT planning solution using a single-center model could efficiently generate plans of comparable quality to manual HNC IMPT plans from centers with differing planning aims. Occasional higher KBP OAR doses highlight the need for beam angle optimization and manual review of KBPs. The solution furthermore demonstrated the potential for robust optimization.

Is accurate contouring of salivary and swallowing structures necessary to spare them in head and neck VMAT plans?

BACKGROUND AND PURPOSE: Current standards for organ-at-risk (OAR) contouring encourage anatomical accuracy which can be resource intensive. Certain OARs may be suitable for alternative delineation strategies. We investigated whether simplified salivary and swallowing structure contouring can still lead to good OAR sparing in automated head and neck cancer (HNC) plans. MATERIALS AND METHODS: For 15 HNC patients, knowledge-based plans (KBPs) using RapidPlan™ were created using: (1) standard clinical contours for all OARs (benchmark-plans), (2) automated knowledge-based contours for the salivary glands, with standard contours for the remaining OARs (SS-plans) and (3) simplified contours (SC-plans) consisting of quick-to-draw tubular structures to account for the oral cavity, salivary glands and swallowing muscles. Individual clinical OAR contours in a RapidPlan™ model were combined to create composite salivary/swallowing structures. These were matched to tube-contours to create SC-plans. All plans were compared based on dose to anatomically accurate clinical OAR contours. RESULTS: Salivary gland delineation in SS-plans required on average 2 min, compared with 7 min for manual delineation of all tubular-contours. Automated atlas-based contours overlapped with, on average, 71% of clinical salivary gland contours while tube-contours overlapped with 95%/75%/93% of salivary gland/oral cavity/swallowing structure contours. On average, SC-plans were comparable to benchmark-plans and SS-plans, with average differences in composite salivary and swallowing structure dose ≤2 Gy and <1 Gy respectively. CONCLUSIONS: Simplified-contours could be created quickly and resulted in clinically acceptable HNC VMAT plans. They can be combined with automated planning to facilitate the implementation of advanced radiotherapy, even when resources are limited.

Evaluation of an Automated Proton Planning Solution.

Purpose Intensity-modulated proton therapy (IMPT) treatments are increasing, however, treatment planning remains complex and prone to variability. RapidPlanTMPT (Varian Medical Systems, Palo Alto, California, USA) is a pre-clinical, proton-specific, automated knowledge-based planning solution which could reduce variability and increase efficiency. It uses a library of previous IMPT treatment plans to generate a model which can predict organ-at-risk (OAR) dose for new patients, and guide IMPT optimization. This study details and evaluates RapidPlanTMPT. Methods IMPT treatment plans for 50 head-and-neck cancer patients populated the model-library. The model was then used to create knowledge-based plans (KBPs) for 10 evaluation-patients. Model quality and accuracy were evaluated using model-provided OAR regression plots and examining the difference between predicted and achieved KBP mean dose. KBP quality was assessed through comparison with respective manual IMPT plans on the basis of boost/elective planning target volume (PTVB/PTVE) homogeneity and OAR sparing. The time to create KBPs was recorded. Results Model quality was good, with an average R2 of 0.85 between dosimetric and geometric features. The model showed high predictive accuracy with differences of <3 Gy between predicted and achieved OAR mean doses for 88/109 OARs. On average, KBPs were comparable to manual IMPT plans with differences of <0.6% in homogeneity. Only 2 of 109 OARs in KBPs had a mean dose >3 Gy more than the manual plan. On average, dose-volume histogram (DVH) predictions required 0.7 minutes while KBP optimization and dose calculation required 4.1 minutes (a 'continue optimization' phase, if required, took an additional 2.8 minutes, on average). Conclusions RapidPlanTMPT demonstrated efficiency and consistency and IMPT KBPs were comparable to manual plans. Because worse OAR sparing in a KBP was not always associated with geometric-outlier warnings, manual plan checks remain important. Such an automated planning solution could also assist in clinical trial quality assurance and overcome the learning curve associated with IMPT.

Using a knowledge-based planning solution to select patients for proton therapy.

BACKGROUND AND PURPOSE: Patient selection for proton therapy by comparing proton/photon treatment plans is time-consuming and prone to bias. RapidPlan™, a knowledge-based-planning solution, uses plan-libraries to model and predict organ-at-risk (OAR) dose-volume-histograms (DVHs). We investigated whether RapidPlan, utilizing an algorithm based only on photon beam characteristics, could generate proton DVH-predictions and whether these could correctly identify patients for proton therapy. MATERIAL AND METHODS: ModelPROT and ModelPHOT comprised 30 head-and-neck cancer proton and photon plans, respectively. Proton and photon knowledge-based-plans (KBPs) were made for ten evaluation-patients. DVH-prediction accuracy was analyzed by comparing predicted-vs-achieved mean OAR doses. KBPs and manual plans were compared using salivary gland and swallowing muscle mean doses. For illustration, patients were selected for protons if predicted ModelPHOT mean dose minus predicted ModelPROT mean dose (ΔPrediction) for combined OARs was ≥6Gy, and benchmarked using achieved KBP doses. RESULTS: Achieved and predicted ModelPROT/ModelPHOT mean dose R2 was 0.95/0.98. Generally, achieved mean dose for ModelPHOT/ModelPROT KBPs was respectively lower/higher than predicted. Comparing ModelPROT/ModelPHOT KBPs with manual plans, salivary and swallowing mean doses increased/decreased by <2Gy, on average. ΔPrediction≥6Gy correctly selected 4 of 5 patients for protons. CONCLUSIONS: Knowledge-based DVH-predictions can provide efficient, patient-specific selection for protons. A proton-specific RapidPlan-solution could improve results.

Detailed evaluation of an automated approach to interactive optimization for volumetric modulated arc therapy plans.

PURPOSE: Interactive optimization during treatment planning requires intermittent adjustment of organ-at-risk (OAR) objectives relative to the dose-volume histogram line. This is a labor-intensive process and the resulting plans are prone to variations in quality. The authors' in-house developed approach to automated interactive optimization (AIO) automatically moves the mouse cursor to adjust the position of on-screen optimization objectives. This allows for the use of more objectives per OAR and results in a more frequent and consistent adjustment of these objectives during optimization. The authors report a detailed evaluation of AIO performance in support of its implementation for routine head and neck cancer (HNC) planning and an evaluation for locally advanced lung cancer (LC) planning which requires a different optimization strategy. METHODS: Volumetric modulated arc therapy AIO plans (APs) were created for 70 HNC patients with a simultaneously integrated boost and 20 LC patients and benchmarked against their respective manually interactively optimized plans (MPs). The same set of optimization objectives and priorities was used for all APs, although planning target volume (PTV) optimization priorities could be increased manually in a subsequent "continue previous optimization" calculation. HNC plans were benchmarked using mean dose to individual and composite OARs and elective/boost PTV (PTVE/PTVB) volumes receiving 95% and 107% of the prescription dose (V95% and V107%, respectively). A clinician performed blinded comparison of 20 APs and respective MPs. LC plans were compared using PTV V95%/V107%, contralateral lung (CL) volume receiving 5 Gy (V5Gy), total lung (TL)-PTV V5Gy/V20Gy, and esophagus and heart V40Gy/V60Gy/mean doses. RESULTS: For HNC, statistically significant improvements in sparing of all OARs, except for the ipsilateral submandibular gland and trachea, were obtained in the APs compared to MPs. Average mean dose to oral cavity, composite salivary, and swallowing structures were 25.4/23.8, 24.2/23.2, and 29.5/25.5 Gy, respectively, for the MPs/APs. PTV heterogeneity was similar: in the APs, PTVB V95% was 0.2% higher while PTV B/PTV E V107% was 0.4%/1.0% lower. In 19 out of 20 HNC patients, the clinician preferred the AP, mainly because of better OAR sparing and PTV dose homogeneity. For LC, APs had a significantly lower CL V5Gy (6.1%), heart mean dose/V60Gy (0.9 Gy/1.2%) and esophagus mean dose/V60Gy (0.9 Gy/2.8%), a nonsignificantly higher TL V20Gy (1.4%), and a slight, but significantly higher dose deposition to the body. PTV dose coverage and homogeneity were similar in the APs and MPs. AIO was considered sufficiently robust for clinical use in LC. CONCLUSIONS: HNC and LC APs were at least as good as, and often of improved quality over MPs. To date, AIO has been clinically implemented for HNC planning.

Effect of Dosimetric Outliers on the Performance of a Commercial Knowledge-Based Planning Solution.

PURPOSE: RapidPlan, a commercial knowledge-based planning solution, uses a model library containing the geometry and associated dosimetry of existing plans. This model predicts achievable dosimetry for prospective patients that can be used to guide plan optimization. However, it is unknown how suboptimal model plans (outliers) influence the predictions or resulting plans. We investigated the effect of, first, removing outliers from the model (cleaning it) and subsequently adding deliberate dosimetric outliers. METHODS AND MATERIALS: Clinical plans from 70 head and neck cancer patients comprised the uncleaned (UC) ModelUC, from which outliers were cleaned (C) to create ModelC. The last 5 to 40 patients of ModelC were replanned with no attempt to spare the salivary glands. These substantial dosimetric outliers were reintroduced to the model in increments of 5, creating Model5 to Model40 (Model5-40). These models were used to create plans for a 10-patient evaluation group. Plans from ModelUC and ModelC, and ModelC and Model5-40 were compared on the basis of boost (B) and elective (E) target volume homogeneity indexes (HIB/HIE) and mean doses to oral cavity, composite salivary glands (compsal) and swallowing (compswal) structures. RESULTS: On average, outlier removal (ModelC vs ModelUC) had minimal effects on HIB/HIE (0%-0.4%) and sparing of organs at risk (mean dose difference to oral cavity and compsal/compswal were ≤0.4 Gy). Model5-10 marginally improved compsal sparing, whereas adding a larger number of outliers (Model20-40) led to deteriorations in compsal up to 3.9 Gy, on average. These increases are modest compared to the 14.9 Gy dose increases in the added outlier plans, due to the placement of optimization objectives below the inferior boundary of the dose-volume histogram-predicted range. CONCLUSIONS: Overall, dosimetric outlier removal from or addition of 5 to 10 outliers to a 70-patient model had marginal effects on resulting plan quality. Although the addition of >20 outliers deteriorated plan quality, the effect was modest. In this study, RapidPlan demonstrated robustness for moderate proportions of salivary gland dosimetric outliers.

Can knowledge-based DVH predictions be used for automated, individualized quality assurance of radiotherapy treatment plans?

BACKGROUND: Treatment plan quality assurance (QA) is important for clinical studies and for institutions aiming to generate near-optimal individualized treatment plans. However, determining how good a given plan is for that particular patient (individualized patient/plan QA, in contrast to running through a checklist of generic QA parameters applied to all patients) is difficult, time consuming and operator-dependent. We therefore evaluated the potential of RapidPlan, a commercial knowledge-based planning solution, to automate this process, by predicting achievable OAR doses for individual patients based on a model library consisting of historical plans with a range of organ-at-risk (OAR) to planning target volume (PTV) geometries and dosimetries. METHODS: A 90-plan RapidPlan model, generated using previously created automatic interactively optimized (AIO) plans, was used to predict achievable OAR dose-volume histograms (DVHs) for the parotid glands, submandibular glands, individual swallowing muscles and oral cavities of 20 head and neck cancer (HNC) patients using a volumetric modulated (RapidArc) simultaneous integrated boost technique. Predicted mean OAR doses were compared with mean doses achieved when RapidPlan was used to make a new plan. Differences between the achieved and predicted DVH-lines were analyzed. Finally, RapidPlan predictions were used to evaluate achieved OAR sparing of AIO and manual interactively optimized plans. RESULTS: For all OARs, strong linear correlations (R(2) = 0.94-0.99) were found between predicted and achieved mean doses. RapidPlan generally overestimated the amount of achievable sparing for OARs with a large degree of OAR-PTV overlap. RapidPlan QA using predicted doses alone identified that for 50 % (10/20) of the manually optimized plans, sparing of the composite salivary glands, oral cavity or composite swallowing muscles could be improved by at least 3 Gy, 5 Gy or 7 Gy, respectively, while this was the case for 20 % (4/20) AIO plans. These predicted gains were validated by replanning the identified patients using RapidPlan. CONCLUSIONS: Strong correlations between predicted and achieved mean doses indicate that RapidPlan could accurately predict achievable mean doses. This shows the feasibility of using RapidPlan DVH prediction alone for automated individualized head and neck plan QA. This has applications in individual centers and clinical trials.

Evaluation of a knowledge-based planning solution for head and neck cancer.

PURPOSE: Automated and knowledge-based planning techniques aim to reduce variations in plan quality. RapidPlan uses a library consisting of different patient plans to make a model that can predict achievable dose-volume histograms (DVHs) for new patients and uses those models for setting optimization objectives. We benchmarked RapidPlan versus clinical plans for 2 patient groups, using 3 different libraries. METHODS AND MATERIALS: Volumetric modulated arc therapy plans of 60 recent head and neck cancer patients that included sparing of the salivary glands, swallowing muscles, and oral cavity were evenly divided between 2 models, Model(30A) and Model(30B), and were combined in a third model, Model60. Knowledge-based plans were created for 2 evaluation groups: evaluation group 1 (EG1), consisting of 15 recent patients, and evaluation group 2 (EG2), consisting of 15 older patients in whom only the salivary glands were spared. RapidPlan results were compared with clinical plans (CP) for boost and/or elective planning target volume homogeneity index, using HI(B)/HI(E) = 100 × (D2% - D98%)/D50%, and mean dose to composite salivary glands, swallowing muscles, and oral cavity (D(sal), D(swal), and D(oc), respectively). RESULTS: For EG1, RapidPlan improved HI(B) and HI(E) values compared with CP by 1.0% to 1.3% and 1.0% to 0.6%, respectively. Comparable D(sal) and D(swal) values were seen in Model(30A), Model(30B), and Model60, decreasing by an average of 0.1, 1.0, and 0.8 Gy and 4.8, 3.7, and 4.4 Gy, respectively. However, differences were noted between individual organs at risk (OARs), with Model(30B) increasing D(oc) by 0.1, 3.2, and 2.8 Gy compared with CP, Model(30A), and Model60. Plan quality was less consistent when the patient was flagged as an outlier. For EG2, RapidPlan decreased D(sal) by 4.1 to 4.9 Gy on average, whereas HI(B) and HI(E) decreased by 1.1% to 1.5% and 2.3% to 1.9%, respectively. CONCLUSIONS: RapidPlan knowledge-based treatment plans were comparable to CP if the patient's OAR-planning target volume geometry was within the range of those included in the models. EG2 results showed that a model including swallowing-muscle and oral-cavity sparing can be applied to patients with only salivary gland sparing. This may allow model library sharing between institutes. Optimal detection of inadequate plans and population of model libraries requires further investigation.