Real-time Cherenkov imaging (CI) visualizes surface dose distribution during external beam radiotherapy (EBRT). Although feasibility has been demonstrated, sustained multi-site implementation with predefined clinical impact metrics remains less well characterized. We evaluated clinic-wide CI integration across two centers over two years and quantified clinical impact, timing of detection, and corrective actions. CI (DoseRT, Vision RT) was incorporated into routine treatment delivery on Varian TrueBeam linear accelerators. All techniques (3D, IMRT, SRS/SBRT) were monitored prospectively. Findings were assigned the highest plan-level clinical impact category: 0-1 (informational/non-actionable), 2 (setup refinement), 3 (workflow/planning modification), and 4 (immediate intervention). During implementation, mandatory physics review of CI images from the first five fractions of new treatment starts was adopted. Outcomes included actionable rate (Categories 2-4), time to first detection, and corrective endpoints (replanning, workflow changes). Among 1,196 plans, 15% were actionable (Category 2: 8%; Category 3: 4%; Category 4: 3%). Twenty-one plans (1.8%) required partial or full replanning and 30 (2.5%) prompted workflow or policy modifications. Approximately half of Category 4 deviations were identified during beam-on and required immediate interruption and correction. Deviations were concentrated in early fractions: 76% were detected by fraction 3 and 89% by fraction 5 (median 2.6). Clinic-wide CI integration is feasible and identifies actionable deviations in 15% of treatment plans, predominantly within the first five fractions. Real-time visualization supported immediate intervention and system-level workflow refinement without measurable disruption to clinical throughput, supporting CI as a complementary verification layer to IGRT/SGRT for patient safety.
Monitoring lung tumors during stereotactic radiotherapy is important for managing breathing motion, yet the high-performance imaging modules on C-arm linear accelerators are mainly used for patient positioning. This study benchmarked a room-mounted against a gantry-mounted kilovolt (kV) imaging device to assess their potential for lung tumor fluoroscopy. The investigated systems were the ExacTrac Dynamic (ETD, Brainlab, Germany) and the X-ray Volumetric Imaging (XVI, Elekta, Sweden). Contrast-to-Noise Ratio (CNR) and entrance air kerma were measured for different imaging settings - tube voltage (70 kV to 130 kV), current (10 mA to 320 mA), exposure time (100 ms and 40 ms) - using two anthropomorphic thorax phantoms housing three tumors. Additionally, CNR change as a function of time, achievable fluoroscopy imaging duration based on X-ray generator and anode heat and X-ray tube cool-down were evaluated. CNRs agreed within 15% for the two systems and changed less than 2% over the investigated fluoroscopy durations. Using default imaging settings and considering the average clinical duration of stereotactic lung treatments, entrance air kermas were 44 mGy (room-mounted) and 41 mGy (gantry-mounted). Achievable fluoroscopy durations with default settings were 87 s and 1350 s, respectively. Anode cool-down times were 125 min (room-mounted) and 215 min (gantry-mounted). The performance metrics of the room-mounted imaging system aligned with those of the gantry-mounted device with regard to CNR and air kerma. The room-mounted system showed a preferable cool-down behavior. However, adaptation of imaging settings is necessary to increase the fluoroscopy duration and avoid overheating.
Brachytherapy is a highly conformal and cost-effective radiotherapy modality, yet its clinical utilization has declined in multiple regions. This study investigated trends in brachytherapy utilization across different Canadian provinces from 2011 to 2020. A national survey was distributed to medical physicists in cancer centres across ten Canadian provinces, collecting data on types of brachytherapy, annual number of treatments, clinical indications, and logistical factors. In Québec, treated clinical indications were obtained through a complementary survey of all radiotherapy centres, extending previously published provincial brachytherapy data. Out of 39 radiotherapy centres in Canada conducting brachytherapy treatments, 25 centres were included in this study. HDR accounted for the majority of treatments (60%-100%) and increased in most provinces, while LDR declined across reporting provinces; PDR comprised ≤ 10% of treatments and was limited to Alberta and Ontario. After adjusting for indication-specific cancer incidence and combining HDR and LDR brachytherapy, utilization trends showed decreases in Alberta and British Columbia, increases in Ontario, Nova Scotia, and Saskatchewan, and no significant change in Manitoba, Québec, or nationally. This study revealed significant regional variations in brachytherapy utilization and clinical practices across Canadian radiotherapy centres. Factors influencing these trends include reimbursement structures, personnel and infrastructure availability, and clinician preferences. The use of HDR modalities is increasing nationally, while LDR is declining. Despite significant regional differences, brachytherapy utilization in Canada was largely sustained over the decade from 2011 to 2020.
Magnetic resonance-guided radiotherapy provides real-time soft-tissue-based targeting during dose delivery. Beyond image guidance, magnetic resonance linear accelerator (MR-Linac) systems offer new opportunities for developing magnetic resonance imaging (MRI)-derived radiomics biomarkers. However, the reproducibility of these biomarkers in a low-field MR-Linac environment remains unclear. This study evaluated the repeatability and reproducibility of MRI-derived radiomic features acquired on a 0.35 T MR-Linac using a tissue-mimicking phantom containing multiple tissue-equivalent materials. Fifty balanced steady-state free precession acquisitions were performed over 10 sessions, enabling both intra- and inter-session analyses. Forty-eight preprocessing pipelines combining normalization, discretization, and bias-field correction were assessed using the coefficient of variation and intraclass correlation coefficient for 68 radiomic features across 12 defined regions of interest. The combination of N4 bias correction and z-score normalization yielded the highest overall stability, with several first-order (e.g. Entropy, Mean) and texture-based (e.g. glcm_DifferenceEntropy and glrlm_ShortRunEmphasis) features showing a coefficient of variation <5% and/or an intraclass correlation coefficient (2,1) > 0.85 across phantom materials. Stability varied with phantom composition, with higher reproducibility in homogeneous PVP-40 and water inserts compared to fat or fibroglandular compartments. Comparison with previous 0.35 T MR-Linac studies identified a consistent subset of robust features, supporting their potential as standardized MRI-derived radiomic biomarkers. These findings demonstrated the critical role of preprocessing and tissue composition in feature reproducibility and highlighted the importance of tissue-mimicking phantoms for validating MR-Linac-based quantitative imaging pipelines.
Quantitative T1 mapping is a major building block in several multiparametric magnetic resonance imaging (MRI) protocols intended for adaptive radiation therapy. The implementation of these protocols is challenging in anatomical sites that experience large physiological motion. The purpose of this study was to implement and validate motion-resolved quantitative T1 mapping on a 1.5 T MRI linear accelerator (MR-Linac) combining non-Cartesian k-space sampling trajectories with compressed sensing (CS) reconstruction. Four 3-dimensional non-Cartesian k-space trajectories were evaluated: kooshball and stack-of-stars sampling using half- and full-spoke coverage. A variable flip angle acquisition was performed using the spoiled gradient-echo sequence. Gradient delay timing was optimized to minimize trajectory-induced artifacts. Eight CS reconstruction strategies were tested using spatial/spatiotemporal regularization operators. Reconstructions were evaluated and sorted by spatial resolution, bias, and variability. Motion-resolved T1 mapping was validated using two standard phantoms, one healthy volunteer, and one kidney cancer patient using respiratory self-gating and phase-sorted reconstruction. All non-Cartesian T1 maps demonstrated high repeatability and low longitudinal bias in phantom studies, with coefficients of variation below 3.3%. Spatiotemporal regularization preserved spatial resolution and quantitative accuracy at undersampling factors up to 20-fold. In human subjects, non-Cartesian T1 mapping provided improved accuracy and reduced variability in mobile abdominal tissues compared to Cartesian acquisitions. Quantitative T1 mapping using non-Cartesian trajectories and CS reconstruction is feasible on a 1.5 T MR-Linac. The proposed approach enables accurate motion-resolved quantitative imaging within clinically practical acquisition times, establishing a foundation for multiparametric MRI in adaptive radiotherapy.
Fused Deposition Modelling (FDM) three-dimensional (3D) printing offers a flexible and economical method for producing radiotherapy phantoms with tailored geometries and material properties. While numerous studies have focused on imaging or photon radiotherapy, research on 3D printing for proton and light ion beam therapy remains limited. Accurate knowledge of the radiological properties of FDM-printed materials is crucial to ensure reliable dose calculation and treatment planning in ion beam therapy. A comprehensive literature review was conducted to identify publications reporting relevant radiological parameters, including mass density, computed tomography (CT) number given in Hounsfield units (HU), electron density, and stopping power, for FDM printing filaments. Based on the collected data, an FDM lookup table was generated, summarising the radiological properties of these materials across different printing settings. A total of 17 material classes comprising 70 distinct filaments were analysed and indexed in an open-access lookup table. Polylactic acid (PLA) was the most frequently investigated material, reported in over 34 publications. Among the investigated radiological parameters, the CT number showed the greatest variability for a given material. For samples printed at 100% infill, values ranged from -180 HU to 227 HU for PLA. Recommendations for reducing this variability through standardised reporting are provided. This review provides an overview of FDM 3D printing materials in ion beam therapy. It serves as a practical reference for clinical personnel, medical physicists, and researchers in selecting suitable materials for radiotherapy applications. Moreover, it highlights the need for standardised characterisation methodologies and 3D printing guidelines.
Auto-segmentation of organs-of-interest (OOI) in cancer patients is essential for facilitating radiotherapy planning and reducing inter-observer variability. Deep learning-based auto-segmentation models have shown promise, but limited transparency and reproducibility hinder their generalizability and clinical acceptability, limiting their use in clinical settings. We introduced an auto-Segmentation Clinical Acceptability & Reproducibility Framework, a comprehensive framework designed to benchmark open-source deep learning models for auto-segmentation of 19 essential OOIs in head and neck cancer (HNC). Reproducibility was achieved through harmonized data curation, standardized model training (training/tune/test of 479/44/59) and assessment workflows. New models can be benchmarked against 12 pre-trained open-source deep learning models, while estimating clinical acceptability using a 5-point Likert scale. The framework codebase is openly available for benchmarking OOI auto-segmentation methods. During development, expert assessment of the best performing model labelled 16/19 AI-generated OOI categories as clinically acceptable with only minor revisions. The framework facilitates benchmarking and expert assessment of AI-driven auto-segmentation tools, addressing the need for transparency and reproducibility in this domain. Through its emphasis on clinical acceptability, our framework fosters the integration of AI models into clinical environments, specifically within radiation therapy.
The American Association of Physicists in Medicine Task Group 307 (AAPM TG-307) recommends validation of electronic portal imaging device (EPID) patient-specific quality assurance (PSQA), including verification of measured and calculated components. This study established a TG-307-aligned commissioning and benchmarking framework for non-transit EPID pre-treatment PSQA. Fraction Zero Absolute Dose (FZAD) was commissioned on two TrueBeam linear accelerators equipped with amorphous silicon (aS) aS1000 and aS1200 EPIDs across flattened and flattening filter-free (FFF) photon beams (6MV/10MV) and two multi-leaf-collimator models. EPID performance was characterised for dose-rate linearity. Independent validation of measured and predicted dose components was performed using ion chamber, phantom-based, and treatment planning system dose planes in geometry replicating EPID configuration. System performance was benchmarked against Portal Dosimetry (PDIP) and phantom-based PSQA across >800 clinical fields. Error-detection sensitivity was evaluated using receiver operating characteristic (ROC) analysis of externally supplied audit plans. The aS1200 panel was linear across dose-rate and source-to-imager distance (SID) conditions, whereas the aS1000 exhibited saturation for FFF beams at shorter SIDs, with up to 5% under-response, with linearity restored at ≥150 cm SID. Independent component validation achieved mean gamma pass rates ≥97% (3%/2 mm). Bland-Altman analysis showed concordance within 95% limits of agreement, supporting equivalence in clinical decision-making. Sensitivity testing demonstrated discrimination at 2%/2 mm (area under ROC curve of 0.886), with improved detection compared to phantom-based PSQA at high specificity. A TG-307-aligned framework enables component-level interpretability and actionable insight for tolerance definition and model optimisation of EPID PSQA.
Accurate delineation of intracranial tumours is crucial for stereotactic radiosurgery (SRS), where target definition directly influences treatment outcome. We developed and clinically integrated an automated multi-tumour segmentation pipeline using three-dimensional nnU-Net models for brain metastases, pituitary adenomas, vestibular schwannomas, and meningiomas. Four independent, tumour-specific models were trained on T1-weighted Magnetisation-Prepared RApid Gradient Echo Magnetic Resonance Imaging data using the nnU‑Net architecture. For model development, 100 cases per tumour-type (n = 400) were used, and to evaluate the clinical workflow, 25 additional cases per tumour-type (n = 100) were processed prospectively. The performance was assessed using the Dice Similarity Coefficient (DSC), the 95th-percentile Hausdorff Distance (HD95), and the Average Symmetric Surface Distance (ASSD). The pipeline continuously monitored incoming Digital Imaging and Communications in Medicine (DICOM) images using a listener and applied the appropriate tumour-specific segmentation model. It, then, automatically exported the DICOM images and the inferred Radiotherapy Structure-Set to the treatment planning system. Among all the tumour-types, vestibular schwannomas achieved the highest performance (DSC: 0.90 ± 0.03; HD95: 0.93 ± 0.34 mm; ASSD: 0.31 ± 0.09 mm) followed by brain metastases (DSC: 0.83 ± 0.08; HD95: 1.33 ± 0.55 mm; ASSD: 0.47 ± 0.19 mm), pituitary adenomas (DSC: 0.81 ± 0.09; HD95: 2.39 ± 1.14 mm; ASSD: 0.78 ± 0.32 mm) and meningiomas (DSC: 0.80 ± 0.11; HD95: 4.46 ± 3.64 mm; ASSD: 1.19 ± 0.80 mm). All tumour-types were segmented with consistent performance (SDDSC < 0.11), and segmentation was completed within two to four minutes per case. The auto-segmentation pipeline enabled consistent and rapid delineation of multiple intracranial tumours, achieving clinically acceptable performance metrics and efficiency suitable for SRS.
This study aimed to characterize the isocentricity of ring-gantry linacs and evaluate variability across institutions and software. Isocentricity is assessed using vendor-provided and user-defined Winston-Lutz (WL) tests. WL was performed at two institutions on a ring-gantry linac using a custom phantom over 1.5 years. Distal and proximal MLCs were tested, with measurements obtained before and after applying 3D shifts. WL analysis was performed using in-house, commercial (C1-C3), and open-source software. Assessed parameters included the 2D ball bearing (BB)-radiation offset, the 3D offset minimizing 2D errors, and 3D isocenter size. 3D isocenter size was additionally computed using gantry-only fields to remove collimator walkout contributions. Results across institutions showed sub-millimeter agreement among software except for C1 (max 0.99 mm difference). Machine-dependent variations were observed (p < 0.03). Proximal and distal banks demonstrated small but statistically significant differences (p < 0.03) in 2D/3D offsets and isocenter size, with greater walkout in the proximal bank and poorer axis alignment in the distal bank. Collimator-walkout-corrected 3D isocenter size was not statistically different between proximal and distal bank or between pre- and post-3D-offset shift. Isocentricity measurements depended on software, MLC bank, pre-/post-offset correction, and institution. For the tested machines, which showed a mean imaging-to-radiation isocenter coincidence of 0.8 mm and a radiation-isocenter size of 0.7 mm, a minimum targeting margin of 1.5 mm would be appropriate to account for geometric uncertainty. With the results from this study, clinicians can choose their own margins based on their machine characteristics.
Plastic materials are widely used as water substitutes in radiotherapy; however, the dosimetric properties of thermoplastic polymers used in 3D printing can vary. A multicentre audit was conducted to quantify variations in geometric accuracy, density and water-mimicking properties of 3D-printed objects. Ten centres printed three polylactic acid (PLA) blocks at varying infills. Block dimensions, including protruding and recessed discs, were measured and the blocks weighed to determine geometric accuracy and mass density. Computed tomography (CT) scans were used to derive the printed infills and mass densities corresponding to water equivalence. Tissue phantom ratios (TPRs) were measured for 6 MV photon beams and compared to vendor-provided reference data for water. A mean error of (-0.2 ± 0.3) mm (mean ± standard deviation, SD) was found for all printed dimensions and disc diameters. Measured CT numbers and mass densities varied by up to 200 Hounsfield units (HU) and 0.20  g/cm3 between centres, respectively. Printed infill and mass densities producing water-equivalent CT density were calculated as (94.5 ± 3.5)% and (1.10 ± 0.03) g/cm3, respectively. Calculated water-equivalent thicknesses formed using combinations of the blocks varied by up to 3.0  mm between centres, with measured TPR data varying by up to 1.2%. Measured and calculated data were normally distributed across centres and fell within ± 2 SD of their respective means. TPR data closely emulated reference values for water. A multicentre audit was completed to develop understanding of geometric and dosimetric errors associated with 3D printing in radiotherapy.
Volumetric Modulated Arc Therapy (VMAT) has enabled highly conformal Total Body Irradiation (TBI), improving dose homogeneity and sparing of organs-of-interest. However, the increased complexity of VMAT-based TBI requires robust quality assurance strategies. This study evaluated the feasibility, accuracy, and reproducibility of Electronic Portal Imaging Device (EPID)-based in-vivo dosimetry (IVD) for VMAT TBI treatments. Forty pediatric patients treated with VMAT-based TBI were retrospectively analyzed with two prescription schemes: 12 Gy/six fractions and 9.99 Gy/three fractions. EPID transmission images were acquired during each VMAT arc and analyzed using a gamma passing rate (Pγ) of 5%/2 mm, 10% threshold. IVD was evaluated across different anatomical regions. The impact of inter-fraction anatomical variations (weight loss, abdominal swelling, gastrointestinal air) was assessed through dose recalculations on modified CT datasets. Phantom measurements were performed to validate the sensitivity of EPID-based IVD to clinically relevant perturbations. IVD measurements showed median Pγ of 97.3 ± 3.2%. Across the forty patients, a mean Pγ of 99.3 ± 0.8% and 99.0 ± 0.4% was observed in the head and thoracic regions respectively, lower values (96.5 ± 3.4% and 94.3 ± 3.6%) were found in the lumbar and pelvic regions respectively. Simulated abdominal swelling ≥1.5 cm reduced target coverage, while weight loss and gastrointestinal air had limited dosimetric impact. Phantom tests confirmed IVD sensitivity to significant anatomical changes. EPID-based IVD was a feasible tool for verifying VMAT-based TBI, enabling detection of clinically relevant inter-fraction anatomical variations and supporting safe treatment delivery.
Fast cone beam computed tomography (CBCT) on ring-gantry systems allows for improved image quality and fast 6-second acquisition. However, 6-second acquisition might pose challenges regarding target delineation and capturing the full range of motion of moving lung tumors. Especially, in patients with slow (period > 6 s) or irregular breathing, capturing the entire tumor motion might not be guaranteed. This study evaluated localization and volumetric accuracy of 6- and 60-second CBCT scans in an in-house dynamic anthropomorphic thorax phantom, with synchronized imaging capabilities. The phantom was scanned for a sinusoidal and four patient-derived breathing patterns, including regular, slow or irregular breathing. Target position and volume for 6- and 60-second acquisitions were compared to ground truth delineation on time-averaged 4D computed tomography (4DCT) reconstruction, assessing if 6-second acquisition is sufficient to accurately capture the tumor motion. For sinusoidal and regular patient-derived motion, both 6- and 60-second CBCT acquisition captured target motion, compared to 4DCT (Dice Similarity Coefficient, DSC > 0.9). For large amplitudes, only one out of three 6-second scans fully captured target motion (DSC > 0.85). For slow and irregular patient-derived patterns, localization errors and volume differences up to 10.3 mm and 119% were observed using 6-second acquisition, compared to 4DCT, with superior localization and volumetric accuracy of the 60-second acquisition. The 6-second protocol showed accurate results, capturing full target motion for regular breathing patterns. Adaptive protocols, taking into account patient-specific breathing periods and irregularities may be preferred in patients exhibiting slow or irregular breathing.
Clinical prognostic models for nasopharyngeal carcinoma (NPC) treated with intensity-modulated radiotherapy (IMRT) with or without chemotherapy remain insufficient to capture tumour heterogeneity. We investigated whether computed tomography (CT)-based signatures add prognostic value for overall survival, progression-free survival, local control and distant control in NPC patients. The study population consisted of 1360 patients with stage I-IVa NPC treated with (chemo)IMRT (2013-2017). Radiomic and deep-learning features were analysed with twelve clinical variables. Radiomic models were built using bootstrap resampling feature selection and multivariable Cox regression; deep-learning models used 3D ResNet-18 or DenseNet-121. Models were evaluated on an internal hold-out test set (n = 409; training set n = 951) with the concordance index and compared against clinical-only reference models. Decision curve analysis was used to assess clinical utility. Adding radiomic primary tumour features (Neighbouring Gray Tone Difference Matrix - coarseness) improved local control concordance index from 0.51 to 0.60 (p = 0.02). A DenseNet-121 combining clinical data with composite primary tumour and lymph node masks achieved the highest distant control (0.68 vs 0.66, p = 0.01). For overall survival and progression-free survival, the improvements were not significant. Decision curve analysis demonstrated net benefit of the DenseNet-121 distant control model over treat-all and treat-none strategies at threshold probabilities of 10-25%. Incorporating CT-based radiomic and deep-learning features into prognostic models significantly improved prediction of local and distant control in NPC, supporting their potential as imaging biomarkers for refined risk stratification.
With advances in deep learning, MRI-only radiotherapy planning (MROP) has been increasingly adopted. However, clinical implementation still requires rigorous evaluation of synthetic CT (sCT) accuracy across multiple anatomical sites.This study aimed to develop and evaluate an all-sites deep learning model for generating sCT from MRI across the head and neck, thorax, abdomen, pelvis, and spine, with comprehensive assessment of image similarity, dose/volume metric fidelity, and patient positioning accuracy. A modified weakly-supervised learning model was trained to generate sCT, incorporating bone-specific and bladder-specific loss term to improve sCT accuracy. Evaluation included: 1. image similarity analysis between sCT and CT using mean absolute error (MAE) and mean error (ME) of Hounsfield Units (HU); 2. dose/volume metric accuracy assessment by recalculating clinical plans on sCT and comparing dose distribution and dose-volume histogram metrics with those calculated on CT; 3. positioning accuracy assessment, including translational and rotational differences, based on rigid CBCT-to-CT versus CBCT-to-sCT registration. The average MAE in HU was lowest in soft-tissue regions and highest in bone. Across all sites, planning target volume mean dose differences between CT- and sCT-based plans were < 1 Gy. Gamma passing rates exceeded 95% under 2 mm/2% criteria for all regions excluding thorax. Positioning accuracy was highest in head&neck and spine, whereas the abdomen and pelvis regions showed greater variability due to soft-tissue motion and MRI-related artifacts. The all-sites sCT model achieved clinically acceptable image quality, dose/volume metric accuracy, and alignment performance across multiple anatomical regions.
Electronic portal imaging device (EPID)-based in-vivo dosimetry (IVD) is widely used to verify radiotherapy dose delivery and detect deviations from anatomical changes, setup uncertainties, or machine-related errors. However, cone-beam computed tomography (CBCT)-guided online-adaptive radiotherapy (oART), which re-optimizes treatment plans on-couch based on daily imaging, lacks a measurement-based quality assurance framework. This study aimed to evaluate the feasibility and accuracy of EPID-based transit IVD for CBCT-based oART. EPID-based IVD data from ten patients receiving short-course oART for rectal cancer on an Ethos linear accelerator were retrospectively analyzed over 50 treatment sessions using a commercial EPID-based IVD framework. Gamma pass rates (GPR) were calculated using global gamma criteria of 3 %/3 mm, 3 %/2 mm, and 2 %/2 mm. Angular dependence and field shape effects were investigated using in-air and in-phantom measurements. Across all patients and fractions, the mean GPR at 3 %/3 mm with a 20 % low-dose threshold was (95.7 ± 1.8) %, with patient-specific means ranging from 91.4 % to 98.0 %. Two sessions showed notable reductions associated with large intra-fractional gas variations. Angular analysis revealed notable GPR variations, particularly for laterally incident fields. EPID-based IVD was a promising quality assurance approach for verifying oART delivery, achieving higher GPRs under tighter evaluation criteria than typically reported for standard image-guided radiotherapy. These findings indicate that workflow-specific tolerance criteria should be defined to reflect the reduced anatomical variation in adaptive treatments, requiring more accurate beam modeling. Automating EPID-based IVD integration into clinical protocols will be essential for safe and efficient implementation.
This study investigated whether scheduled adaptive radiotherapy (ART) improved delivered dose to organs at risk (OAR) in patients with locally advanced head and neck cancer treated with dose painting (DP). Delivered doses were estimated for 81 patients who were prospectively treated with DP + ART using deformable image registration between CBCT and the planning CT's. Three simulations were conducted, evaluating (1) a scenario without replanning (SimnoART), (2) a scenario with ART at fraction 12 (SimART), and (3) the clinical practice with ad hoc replanning (Simdelivered). It was further evaluated whether selecting patients for ART using accumulated dose in the first 10 fractions (Df10) to the parotid glands and larynx would improve ART efficacy. In SimnoART, 41% of patients had delivered dose deviations ≥ 3 Gy compared to the planned dose in any OAR, primarily in the parotid glands and larynx. No significant differences were seen between SimnoART, SimART and Simdelivered (P ≥ 0.10). Df10 predicted relevant changes upon completing treatment with AUC ≥ 0.95. By selecting patients for ART using Df10, delivered dose significantly improved for the larynx (P ≤ 0.01). Although relevant dose differences between planned and delivered doses were seen in almost half of the patients without ART, minimal improvements in delivered doses were seen introducing ART. Nonetheless, Df10 was prognostic for relevant changes upon completing treatment and selecting patients for ART significantly improved larynx dose. Incorporating accumulated dose into patient selection for ART could help avoid unforeseen increases in delivered dose.
Manual contouring of organs of interest (OoIs) is a major bottleneck in pancreatic magnetic resonance-guided online adaptive radiotherapy (oART). We developed C-SegDeform, a data-efficient conditional segmentation framework that used structure-guided deformation-based augmentations to simulate plausible inter-fraction anatomical variation and leveraged organ-specific conditioning as an alternative to registration-based contour propagation (Prop-ROIs) in limited-data settings. Forty balanced 3DVane images from 12 patients were manually contoured and pre-processed, including duodenum, both kidneys, liver, large and small bowel, spinal canal, spleen, and stomach. The training dataset (26 images) was augmented by simulating plausible session images via structure-guided deformations. Data was arranged for conditional segmentation and leave-one-out cross-validation using the nnU-Net framework. Analysis included geometric (DSC, average surface distance (ASD), and 95th percentile of Hausdorff distance (HD95); against Prop-ROIs and TotalSegmentator MRI), dose (via 8 treatment plans, measuring D 0 . 1 c m 3 and D 50 % discrepancies relative to prescribed dose), and clinical (Likert scale and post-auto-contour editing times by two oncology consultants) assessments. C-SegDeform (DSC: 0 . 88 ± 0 . 10 , ASD: 2 . 9 ± 2 . 4 mm , HD95: 10 . 5 ± 8 . 9 mm ) outperformed Prop-ROIs ( 0 . 70 ± 0 . 17 , 8 . 5 ± 9 . 9 mm , 16 . 7 ± 9 . 8 mm ) and TotalSegmentator ( 0 . 75 ± 0 . 16 , 5 . 1 ± 3 . 4 mm , 14 . 3 ± 10 . 6 mm ), and showed smaller relative dose deviations from ground-truth contours ( D 0 . 1 c m 3 : 2 . 2 % ± 2 . 5 % , D 50 % : 0 . 5 % ± 0 . 9 % ) than Prop-ROIs ( 6 . 8 % ± 5 . 7 % , 3 . 3 % ± 4 . 0 % ). Clinicians ranked 98% of C-SegDeform contours as requiring no or minor edits (vs. 70% for Prop-ROIs), with editing times under 6 min and full OoI generation completed in under 40 s. C-SegDeform provided a robust alternative to contour propagation, generating accurate contours rapidly with minimal edits, thereby reducing on-couch time and streamlining the workflow.
Fully automated workflows integrating deep learning-based segmentation and treatment planning have shown promising results in research settings but require prospective evaluation under routine clinical conditions. This study assessed the clinical usability of a fully automated workflow for prostate radiotherapy compared with the conventional clinical workflow. Twenty-two consecutively treated patients undergoing prostate radiotherapy were included. For each patient, treatment planning was first performed according to the conventional workflow, followed by execution of a fully automated workflow combining deep learning-based segmentation and planning models within a commercial treatment planning system. Plans and structure sets were evaluated in a blinded, standardized manner by board-certified radiation oncologists and rated as "good", "acceptable", or "unacceptable". The preferred plan was selected for treatment. Processing times were recorded. Clinically acceptable results were achieved in 86% of fully automated cases compared with 95% in the conventional workflow. Conventional plans were rated more frequently as "good" (17/22 vs 12/22), mainly due to differences in target delineation. Fully automated clinical target volumes were consistently smaller (median 54.2 cm3 vs 68.3 cm3; p < 0.001). The fully automated workflow required a median of 9 min (range 8-10) compared with 87 min (range 70-130) for the conventional workflow. A fully automated workflow for prostate radiotherapy substantially reduces workload and processing time while achieving a high rate of clinical acceptability. Implementation as a draft-first approach with mandatory physician review appears feasible, although refinement of automated target delineation is warranted.
Although pencil beam scanning provides superior dose conformity, delivery uncertainties remain in beam monitoring and steering systems. This study evaluates a machine learning model that predicts delivered spot positions from plan parameters as a complementary quality assurance support tool. The dataset consisted of a single routine quality assurance treatment plan and 64 corresponding delivery log files from a Hitachi proton scanning beam system (32 files each from gantry 1 and 2), collected over three months across various proton energies. Statistical analyses, including Levene's test, Welch's analysis of variance, and Games-Howell post hoc tests, assessed the effects of beam energy, treatment day, and room on spot position errors. Machine learning regression models were trained using 48 delivery log files to predict delivered spot positions from treatment planning data. Model performance was assessed using mean squared errors, R2 score, and Euclidean distance. Machine learning models demonstrated feasibility of spot position prediction with high accuracy (R2 = 0.999; mean squared errors of 0.021 mm2 and 0.003 mm2), with Euclidean errors <0.13 mm. Predicted dose distributions closely matched planned distributions, with an average absolute mean dose difference of 2.34 × 10-5 Gy, and a maximum absolute voxel-wise dose difference of 1.796 Gy, defined as the largest absolute difference between predicted and planned dose values at corresponding voxels. This feasibility study demonstrates that machine-learning-based prediction of delivered spot positions can achieve sub-millimeter accuracy, potentially enhancing the precision and reliability of quality assurance processes in proton therapy.