Framework for Radiation Oncology Department-wide Evaluation and Implementation of Commercial Artificial Intelligence Autocontouring.

PURPOSE: Artificial intelligence (AI)-based autocontouring in radiation oncology has potential benefits such as standardization and time savings. However, commercial AI solutions require careful evaluation before clinical integration. We developed a multidimensional evaluation method to test pretrained AI-based automated contouring solutions across a network of clinics. METHODS AND MATERIALS: Curated data included 121 patient planning computed tomography (CT) scans with a total of 859 clinically approved contours used for treatment from 4 clinics. Regions of interest (ROIs) were generated with 3 commercial AI-based automated contouring software solutions (AI1, AI2, AI3) spanning the following disease sites: brain, head and neck (H&N), thorax, abdomen, and pelvis. Quantitative agreement between AI-generated and clinical contours was measured by Dice similarity coefficient (DSC) and Hausdorff distance (HD). Qualitative assessment was performed by multiple experts scoring blinded AI-contours using a Likert scale. Workflow and usability surveying was also conducted. RESULTS: AI1, AI2, and AI3 contours had high quantitative agreement in 27.8%, 32.8%, and 34.1% of cases (DSC >0.9), performing well in pelvis (median DSC = 0.86/0.88/0.91) and thorax (median DSC = 0.91/0.89/0.91). All 3 solutions had low quantitative agreement in 7.4%, 8.8%, and 6.1% of cases (DSC <0.5), performing worse in brain (median DSC = 0.65/0.78/0.75) and H&N (median DSC = 0.76/0.80/0.81). Qualitatively, AI1 and AI2 contours were acceptable (rated 1-2) with at most minor edits in 70.7% and 74.6% of ROIs (2906 ratings), higher for abdomen (AI1: 79.2%) and thorax (AI2: 90.2%), and lower for H&N (29.0/35.6%). An end-user survey showed strong user preference for full automation and mixed preferences for accuracy versus total number of structures generated. CONCLUSIONS: Our evaluation method provided a comprehensive analysis of both quantitative and qualitative measures of commercially available pretrained AI autocontouring algorithms. The evaluation framework served as a roadmap for clinical integration that aligned with user workflow preference.

Automated treatment planning for proton pencil beam scanning using deep learning dose prediction and dose-mimicking optimization.

PURPOSE: The purpose of this study is to investigate the use of a deep learning architecture for automated treatment planning for proton pencil beam scanning (PBS). METHODS: A 3-dimensional (3D) U-Net model has been implemented in a commercial treatment planning system (TPS) that uses contoured regions of interest (ROI) binary masks as model inputs with a predicted dose distribution as the model output. Predicted dose distributions were converted to deliverable PBS treatment plans using a voxel-wise robust dose mimicking optimization algorithm. This model was leveraged to generate machine learning (ML) optimized plans for patients receiving proton PBS irradiation of the chest wall. Model training was carried out on a retrospective set of 48 previously-treated chest wall patient treatment plans. Model evaluation was carried out by generating ML-optimized plans on a hold-out set of 12 contoured chest wall patient CT datasets from previously treated patients. Clinical goal criteria and gamma analysis were used to compare dose distributions of the ML-optimized plans against the clinically approved plans across the test patients. RESULTS: Statistical analysis of mean clinical goal criteria indicates that compared to the clinical plans, the ML optimization workflow generated robust plans with similar dose to the heart, lungs, and esophagus while achieving superior dosimetric coverage to the PTV chest wall (clinical mean V95 = 97.6% vs. ML mean V95 = 99.1%, p < 0.001) across the 12 test patients. CONCLUSIONS: ML-based automated treatment plan optimization using the 3D U-Net model can generate treatment plans of similar clinical quality compared to human-driven optimization.

Improved lateral penumbra for proton ocular treatments on a general-purpose spot scanning beamline.

PURPOSE: An ocular applicator that fits a commercial proton snout with an upstream range shifter to allow for treatments with sharp lateral penumbra is described. MATERIALS AND METHODS: The validation of the ocular applicator consisted of a comparison of range, depth doses (Bragg peaks and spread out Bragg peaks), point doses, and 2-D lateral profiles. Measurements were made for three field sizes, 1.5, 2, and 3 cm, resulting in 15 beams. Distal and lateral penumbras were simulated in the treatment planning system for seven range-modulation combinations for beams typical of ocular treatments and a field size of 1.5 cm, and penumbra values were compared to published literature. RESULTS: All the range errors were within 0.5 mm. The maximum averaged local dose differences for Bragg peaks and SOBPs were 2.6% and 1.1%, respectively. All the 30 measured point doses were within +/-3% of the calculated. The measured lateral profiles, analyzed through gamma index analysis and compared to the simulated, had pass rates greater than 96% for all the planes. The lateral penumbra increased linearly with depth, from 1.4 mm at 1 cm depth to 2.5 mm at 4 cm depth. The distal penumbra ranged from 3.6 to 4.4 mm and increased linearly with the range. The treatment time for a single 10 Gy (RBE) fractional dose ranged from 30 to 120 s, depending on the shape and size of the target. CONCLUSIONS: The ocular applicator's modified design allows lateral penumbra similar to dedicated ocular beamlines while enabling planners to use modern treatment tools such as Monte Carlo and full CT-based planning with increased flexibility in beam placement.

Multitask Learning Radiomics on Longitudinal Imaging to Predict Survival Outcomes following Risk-Adaptive Chemoradiation for Non-Small Cell Lung Cancer.

Medical imaging provides quantitative and spatial information to evaluate treatment response in the management of patients with non-small cell lung cancer (NSCLC). High throughput extraction of radiomic features on these images can potentially phenotype tumors non-invasively and support risk stratification based on survival outcome prediction. The prognostic value of radiomics from different imaging modalities and time points prior to and during chemoradiation therapy of NSCLC, relative to conventional imaging biomarker or delta radiomics models, remains uncharacterized. We investigated the utility of multitask learning of multi-time point radiomic features, as opposed to single-task learning, for improving survival outcome prediction relative to conventional clinical imaging feature model benchmarks. Survival outcomes were prospectively collected for 45 patients with unresectable NSCLC enrolled on the FLARE-RT phase II trial of risk-adaptive chemoradiation and optional consolidation PD-L1 checkpoint blockade (NCT02773238). FDG-PET, CT, and perfusion SPECT imaging pretreatment and week 3 mid-treatment was performed and 110 IBSI-compliant pyradiomics shape-/intensity-/texture-based features from the metabolic tumor volume were extracted. Outcome modeling consisted of a fused Laplacian sparse group LASSO with component-wise gradient boosting survival regression in a multitask learning framework. Testing performance under stratified 10-fold cross-validation was evaluated for multitask learning radiomics of different imaging modalities and time points. Multitask learning models were benchmarked against conventional clinical imaging and delta radiomics models and evaluated with the concordance index (c-index) and index of prediction accuracy (IPA). FDG-PET radiomics had higher prognostic value for overall survival in test folds (c-index 0.71 [0.67, 0.75]) than CT radiomics (c-index 0.64 [0.60, 0.71]) or perfusion SPECT radiomics (c-index 0.60 [0.57, 0.63]). Multitask learning of pre-/mid-treatment FDG-PET radiomics (c-index 0.71 [0.67, 0.75]) outperformed benchmark clinical imaging (c-index 0.65 [0.59, 0.71]) and FDG-PET delta radiomics (c-index 0.52 [0.48, 0.58]) models. Similarly, the IPA for multitask learning FDG-PET radiomics (30%) was higher than clinical imaging (26%) and delta radiomics (15%) models. Radiomics models performed consistently under different voxel resampling conditions. Multitask learning radiomics for outcome modeling provides a clinical decision support platform that leverages longitudinal imaging information. This framework can reveal the relative importance of different imaging modalities and time points when designing risk-adaptive cancer treatment strategies.

Treatment of ocular tumors through a novel applicator on a conventional proton pencil beam scanning beamline.

Treatment of ocular tumors on dedicated scattering-based proton therapy systems is standard afforded due to sharp lateral and distal penumbras. However, most newer proton therapy centers provide pencil beam scanning treatments. In this paper, we present a pencil beam scanning (PBS)-based ocular treatment solution. The design, commissioning, and validation of an applicator mount for a conventional PBS snout to allow for ocular treatments are given. In contrast to scattering techniques, PBS-based ocular therapy allows for inverse planning, providing planners with additional flexibility to shape the radiation field, potentially sparing healthy tissues. PBS enables the use of commercial Monte Carlo algorithms resulting in accurate dose calculations in the presence of heterogeneities and fiducials. The validation consisted of small field dosimetry measurements of point doses, depth doses, and lateral profiles relevant to ocular therapy. A comparison of beam properties achieved through the applicator against published literature is presented. We successfully showed the feasibility of PBS-based ocular treatments.

Validation and practical implementation of seated position radiotherapy in a commercial TPS for proton therapy.

PURPOSE: This work aims to validate new 6D couch features and their implementation for seated radiotherapy in RayStation (RS) treatment planning system (TPS). MATERIALS AND METHODS: In RS TPS, new 6D couch features are (i) chair support device, (ii) patient treatment option of "Sitting: face towards the front of the chair", and (iii) patient support pitch and roll capabilities. The validation of pitch and roll was performed by comparing TPS generated DRRs with planar x-rays. Dosimetric tests through measurement by 2D ion chamber array were performed for beams created with varied scanning and treatment orientation and 6D couch rotations. For the implementation of 6D couch features for treatments in a seated position, the TPS and oncology information system (Mosaiq) settings are described for a commercial chair. An end-to-end test using an anthropomorphic phantom was performed to test the complete workflow from simulation to treatment delivery. RESULTS: The 6D couch features were found to have a consistent implementation that met IEC 61712 standard. The DRRs were found to have an acceptable agreement with planar x-rays based on visual inspection. For dose map comparison between measured and calculated, the gamma index analysis for all the beams was >95% at a 3% dose-difference and 3 mm distance-to-agreement tolerances. For an end-to end-testing, the phantom was successfully set up at isocenter in the seated position and treatment was delivered. CONCLUSIONS: Chair-based treatments in a seated position can be implemented in RayStation through the use of newly released 6D couch features.

A machine learning-based framework for delivery error prediction in proton pencil beam scanning using irradiation log-files.

PURPOSE: This study aims to investigate the use of machine learning models for delivery error prediction in proton pencil beam scanning (PBS) delivery. METHODS: A dataset of planned and delivered PBS spot parameters was generated from a set of 20 prostate patient treatments. Planned spot parameters (spot position, MU and energy) were extracted from the treatment planning system (TPS) for each beam. Delivered spot parameters were extracted from irradiation log-files for each beam delivery following treatment. The dataset was used as a training dataset for three machine learning models which were trained to predict delivered spot parameters based on planned parameters. K-fold cross validation was employed for hyper-parameter tuning and model selection where the mean absolute error (MAE) was used as the model evaluation metric. The model with lowest MAE was then selected to generate a predicted dose distribution for a test prostate patient within a commercial TPS. RESULTS: Analysis of the spot position delivery error between planned and delivered values resulted in standard deviations of 0.39 mm and 0.44 mm for x and y spot positions respectively. Prediction error standard deviation values of spot positions using the selected model were 0.22 mm and 0.11 mm for x and y spot positions respectively. Finally, a three-way comparison of dose distributions and DVH values for select OARs indicates that the random-forest-predicted dose distribution within the test prostate patient was in closer agreement to the delivered dose distribution than the planned distribution. CONCLUSIONS: PBS delivery error can be accurately predicted using machine learning techniques.

Advanced proton beam dosimetry part II: Monte Carlo vs. pencil beam-based planning for lung cancer.

BACKGROUND: Proton pencil beam (PB) dose calculation algorithms have limited accuracy within heterogeneous tissues of lung cancer patients, which may be addressed by modern commercial Monte Carlo (MC) algorithms. We investigated clinical pencil beam scanning (PBS) dose differences between PB and MC-based treatment planning for lung cancer patients. METHODS: With IRB approval, a comparative dosimetric analysis between RayStation MC and PB dose engines was performed on ten patient plans. PBS gantry plans were generated using single-field optimization technique to maintain target coverage under range and setup uncertainties. Dose differences between PB-optimized (PBopt), MC-recalculated (MCrecalc), and MC-optimized (MCopt) plans were recorded for the following region-of-interest metrics: clinical target volume (CTV) V95, CTV homogeneity index (HI), total lung V20, total lung VRX (relative lung volume receiving prescribed dose or higher), and global maximum dose. The impact of PB-based and MC-based planning on robustness to systematic perturbation of range (±3% density) and setup (±3 mm isotropic) was assessed. Pairwise differences in dose parameters were evaluated through non-parametric Friedman and Wilcoxon sign-rank testing. RESULTS: In this ten-patient sample, CTV V95 decreased significantly from 99-100% for PBopt to 77-94% for MCrecalc and recovered to 99-100% for MCopt (P<10-5). The median CTV HI (D95/D5) decreased from 0.98 for PBopt to 0.91 for MCrecalc and increased to 0.95 for MCopt (P<10-3). CTV D95 robustness to range and setup errors improved under MCopt (ΔD95 =-1%) compared to MCrecalc (ΔD95 =-6%, P=0.006). No changes in lung dosimetry were observed for large volumes receiving low to intermediate doses (e.g., V20), while differences between PB-based and MC-based planning were noted for small volumes receiving high doses (e.g., VRX). Global maximum patient dose increased from 106% for PBopt to 109% for MCrecalc and 112% for MCopt (P<10-3). CONCLUSIONS: MC dosimetry revealed a reduction in target dose coverage under PB-based planning that was regained under MC-based planning along with improved plan robustness. MC-based optimization and dose calculation should be integrated into clinical planning workflows of lung cancer patients receiving actively scanned proton therapy.

Advanced Proton Beam Dosimetry Part I: review and performance evaluation of dose calculation algorithms.

The accuracy of dose calculation is vital to the quality of care for patients undergoing proton beam therapy (PBT). Currently, the dose calculation algorithms available in commercial treatment planning systems (TPS) in PBT are classified into two classes: pencil beam (PB) and Monte-Carlo (MC) algorithms. PB algorithms are still regarded as the standard of practice in PBT, but they are analytical approximations whereas MC algorithms use random sampling of interaction cross-sections that represent the underlying physics to simulate individual particles trajectories. This article provides a brief review of PB and MC dose calculation algorithms employed in commercial treatment planning systems and their performance comparison in phantoms through simulations and measurements. Deficiencies of PB algorithms are first highlighted by a simplified simulation demonstrating the transport of a single sub-spot of proton beam that is incident at an oblique angle in a water phantom. Next, more typical cases of clinical beams in water phantom are presented and compared to measurements. The inability of PB to correctly predict the range and subsequently distal fall-off is emphasized. Through the presented examples, it is shown how dose errors as high as 30% can result with use of a PB algorithm. These dose errors can be minimized to clinically acceptable levels of less than 5%, if MC algorithm is employed in TPS. As a final illustration, comparison between PB and MC algorithm is made for a clinical beam that is use to deliver uniform dose to a target in a lung section of an anthropomorphic phantom. It is shown that MC algorithm is able to correctly predict the dose at all depths and matched with measurements. For PB algorithm, there is an increasing mismatch with the measured doses with increasing tissue heterogeneity. The findings of this article provide a foundation for the second article of this series to compare MC vs. PB based lung cancer treatment planning.

Dosimetric evaluation of a commercial proton spot scanning Monte-Carlo dose algorithm: comparisons against measurements and simulations.

RaySearch Americas Inc. (NY) has introduced a commercial Monte Carlo dose algorithm (RS-MC) for routine clinical use in proton spot scanning. In this report, we provide a validation of this algorithm against phantom measurements and simulations in the GATE software package. We also compared the performance of the RayStation analytical algorithm (RS-PBA) against the RS-MC algorithm. A beam model (G-MC) for a spot scanning gantry at our proton center was implemented in the GATE software package. The model was validated against measurements in a water phantom and was used for benchmarking the RS-MC. Validation of the RS-MC was performed in a water phantom by measuring depth doses and profiles for three spread-out Bragg peak (SOBP) beams with normal incidence, an SOBP with oblique incidence, and an SOBP with a range shifter and large air gap. The RS-MC was also validated against measurements and simulations in heterogeneous phantoms created by placing lung or bone slabs in a water phantom. Lateral dose profiles near the distal end of the beam were measured with a microDiamond detector and compared to the G-MC simulations, RS-MC and RS-PBA. Finally, the RS-MC and RS-PBA were validated against measured dose distributions in an Alderson-Rando (AR) phantom. Measurements were made using Gafchromic film in the AR phantom and compared to doses using the RS-PBA and RS-MC algorithms. For SOBP depth doses in a water phantom, all three algorithms matched the measurements to within ±3% at all points and a range within 1 mm. The RS-PBA algorithm showed up to a 10% difference in dose at the entrance for the beam with a range shifter and >30 cm air gap, while the RS-MC and G-MC were always within 3% of the measurement. For an oblique beam incident at 45°, the RS-PBA algorithm showed up to 6% local dose differences and broadening of distal fall-off by 5 mm. Both the RS-MC and G-MC accurately predicted the depth dose to within ±3% and distal fall-off to within 2 mm. In an anthropomorphic phantom, the gamma index (dose tolerance = 3%, distance-to-agreement = 3 mm) was greater than 90% for six out of seven planes using the RS-MC, and three out seven for the RS-PBA. The RS-MC algorithm demonstrated improved dosimetric accuracy over the RS-PBA in the presence of homogenous, heterogeneous and anthropomorphic phantoms. The computation performance of the RS-MC was similar to the RS-PBA algorithm. For complex disease sites like breast, head and neck, and lung cancer, the RS-MC algorithm will provide significantly more accurate treatment planning.

Dosimetric Comparison of Pencil-Beam Scanning and Photon-Based Radiation Therapy as a Boost in Carcinoma of Cervix.

PURPOSE: Brachytherapy is essential for local treatment in cervical carcinoma, but some patients are not suitable for it. Presently, for these patients, the authors prefer a boost by using intensity-modulated radiation therapy (IMRT). The authors evaluated the dosimetric comparison of proton-modulated radiation therapy versus IMRT and volumetric-modulated arc therapy (VMAT) as a boost to know whether protons can replace photons. PATIENTS AND METHODS: Five patients who received external beam radiation therapy to the pelvis by IMRT were reviewed. Three different plans were made, including pencil beam scanning (PBS), IMRT, and VMAT. The prescribed planning target volume (PTV) was 20 Gy in 4 fractions. The dose to 95% PTV (D95%), the conformity index, and the homogeneity index were evaluated for PTV. The Dmax, D2cc, and Dmean were evaluated for organs at risk along with the integral dose of normal tissue and organs at risk. RESULTS: The PTV coverage was optimal and homogeneous with modulated protons and photons. For PBS, coverage D95% was 20.01 ± 0.02 Gy (IMRT, 20.08 ± 0.06 Gy; VMAT, 20.1 ± 0.04 Gy). For the organs at risk, Dmax of the bladder for PBS was 21.05 ± 0.05 Gy (IMRT, 20.8 ± 0.21 Gy; VMAT, 21.65 ± 0.41 Gy) while the Dmax for the rectum for PBS was 21.04 ± 0.03 Gy (IMRT, 20.81 ± 0.12 Gy; VMAT, 21.66 ± 0.38 Gy). Integral dose to normal tissues in PBS was 14.17 ± 2.65 Gy (IMRT, 25.29 ± 6.35 Gy; VMAT, 25.24 ± 6.24 Gy). CONCLUSIONS: Compared with photons, modulated protons provide comparable conformal plans. However, PBS reduces the integral dose to critical structures significantly compared with IMRT and VMAT. Although PBS may be a better alternative for such cases, further research is required to substantiate such findings.