Search

PurposeThe purpose of this work was to develop a knowledge-based dose prediction system using a convolution neural network (CNN) for cervical brachytherapy treatments with a tandem-and-ovoid applicator.MethodsA 3D U-NET CNN was utilized to make voxel-wise dose predictions based on organ-at-risk (OAR), high-risk clinical target volume (HRCTV), and possible source location geometry. The model comprised 395 previously treated cases: training (273), validation (61), test (61). To assess voxel prediction accuracy, we evaluated dose differences in all cohorts across the dose range of 20-130% of prescription, mean (SD) and standard deviation (σ), as well as isodose dice similarity coefficients for clinical and/or predicted dose distributions. We examined discrete Dose-Volume Histogram (DVH) metrics utilized for brachytherapy plan quality assessment (HRCTV D90%; bladder, rectum, and sigmoid D2cc) with ΔDx=Dx,actual-Dx,predicted mean, standard deviation, and Pearson correlation coefficient further quantifying model performance.ResultsRanges of voxel-wise dose difference accuracy (δD¯±σ) for 20-130% dose interval in training (test) sets ranged from [-0.5% ± 2.0% to +2.0% ± 14.0%] ([-0.1% ± 4.0% to +4.0% ± 26.0%]) in all voxels, [-1.7% ± 5.1% to -3.5% ± 12.8%] ([-2.9% ± 4.8% to -2.6% ± 18.9%]) in HRCTV, [-0.02% ± 2.40% to +3.2% ± 12.0%] ([-2.5% ± 3.6% to +0.8% ± 12.7%]) in bladder, [-0.7% ± 2.4% to +15.5% ± 11.0%] ([-0.9% ± 3.2% to +27.8% ± 11.6%]) in rectum, and [-0.7% ± 2.3% to +10.7% ± 15.0%] ([-0.4% ± 3.0% to +18.4% ± 11.4%]) in sigmoid. Isodose dice similarity coefficients ranged from [0.96,0.91] for training and [0.94,0.87] for test cohorts. Relative DVH metric prediction in the training (test) set were HRCTV ΔD¯90±σΔD = -0.19 ± 0.55Gy (-0.09 ± 0.67 Gy), bladder ΔD¯2cc±σΔD = -0.06 ± 0.54Gy (-0.17 ± 0.67 Gy), rectum ΔD¯2cc±σΔD= -0.03 ± 0.36Gy (-0.04 ± 0.46 Gy), and sigmoid ΔD¯2cc±σΔD = -0.01 ± 0.34Gy (0.00 ± 0.44 Gy).ConclusionsA 3D knowledge-based dose predictions provide voxel-level and DVH metric estimates that could be used for treatment plan quality control and data-driven plan guidance.

PurposeCurrently, there is a lack of patient-specific tools to guide brachytherapy planning and applicator choice for cervical cancer. The purpose of this study is to evaluate the accuracy of organ-at-risk (OAR) dose predictions using knowledge-based intracavitary models, and the use of these models and clinical data to determine the dosimetric differences of tandem-and-ring (T&R) and tandem-and-ovoids (T&O) applicators.Materials and methodsKnowledge-based models, which predict organ D2cc, were trained on 77/75 cases and validated on 32/38 for T&R/T&O applicators. Model performance was quantified using ΔD2cc=D2cc,actual-D2cc,predicted, with standard deviation (σ(ΔD2cc)) representing precision. Model-predicted applicator dose differences were determined by applying T&O models to T&R cases, and vice versa, and compared to clinically-achieved D2cc differences. Applicator differences were assessed using a Student's t-test (p < 0.05 significant).ResultsValidation T&O/T&R model precision was 0.65/0.55 Gy, 0.55/0.38 Gy, and 0.43/0.60 Gy for bladder, rectum and sigmoid, respectively, and similar to training. When applying T&O/T&R models to T&R/T&O cases, bladder, rectum and sigmoid D2cc values in EQD2 were on average 5.69/2.62 Gy, 7.31/6.15 Gy and 3.65/0.69 Gy lower for T&R, with similar HRCTV volume and coverage. Clinical data also showed lower T&R OAR doses, with mean EQD2 D2cc deviations of 0.61 Gy, 7.96 Gy (p < 0.01) and 5.86 Gy (p < 0.01) for bladder, rectum and sigmoid.ConclusionsAccurate knowledge-based dose prediction models were developed for two common intracavitary applicators. These models could be beneficial for standardizing and improving the quality of brachytherapy plans. Both models and clinical data suggest that significant OAR sparing can be achieved with T&R over T&O applicators, particularly for the rectum.

BackgroundWhole-brain radiotherapy (WBRT) remains an important treatment for over 200,000 cancer patients in the United States annually. Hippocampal-avoidant WBRT (HA-WBRT) reduces neurocognitive toxicity compared to standard WBRT, but HA-WBRT contouring and planning are more complex and time-consuming than standard WBRT. We designed and evaluated a workflow using commercially available artificial intelligence tools for automated hippocampal segmentation and treatment planning to efficiently generate clinically acceptable HA-WBRT radiotherapy plans.MethodsWe retrospectively identified 100 consecutive adult patients treated for brain metastases outside the hippocampal region. Each patient's T1 post-contrast brain MRI was processed using NeuroQuant, an FDA-approved software that provides segmentations of brain structures in less than 8 min. Automated hippocampal segmentations were reviewed for accuracy, then converted to files compatible with a commercial treatment planning system, where hippocampal avoidance regions and planning target volumes (PTV) were generated. Other organs-at-risk (OARs) were previously contoured per clinical routine. A RapidPlan knowledge-based planning routine was applied for a prescription of 30 Gy in 10 fractions using volumetric modulated arc therapy (VMAT) delivery. Plans were evaluated based on NRG CC001 dose-volume objectives (Brown et al. in J Clin Oncol, 2020).ResultsOf the 100 cases, 99 (99%) had acceptable automated hippocampi segmentations without manual intervention. Knowledge-based planning was applied to all cases; the median processing time was 9 min 59 s (range 6:53-13:31). All plans met per-protocol dose-volume objectives for PTV per the NRG CC001 protocol. For comparison, only 65.5% of plans on NRG CC001 met PTV goals per protocol, with 26.1% within acceptable variation. In this study, 43 plans (43%) met OAR constraints, and the remaining 57 (57%) were within acceptable variation, compared to 42.5% and 48.3% on NRG CC001, respectively. No plans in this study had unacceptable dose to OARs, compared to 0.8% of manually generated plans from NRG CC001. 8.4% of plans from NRG CC001 were not scored or unable to be evaluated.ConclusionsAn automated pipeline harnessing the efficiency of commercially available artificial intelligence tools can generate clinically acceptable VMAT HA-WBRT plans with minimal manual intervention. This process could improve clinical efficiency for a treatment established to improve patient outcomes over standard WBRT.

PurposeThe purpose of this study is to explore knowledge-based organ-at-risk dose estimation for intracavitary brachytherapy planning for cervical cancer. Using established external-beam knowledge-based dose-volume histogram (DVH) estimation methods, we sought to predict bladder, rectum, and sigmoid D2cc for tandem and ovoid treatments.Methods and materialsA total of 136 patients with loco-regionally advanced cervical cancer treated with 456 (356:100 training:validation ratio) CT-based tandem and ovoid brachytherapy fractions were analyzed. Single fraction prescription doses were 5.5-8 Gy with dose criteria for the high-risk clinical target volume, bladder, rectum, and sigmoid. DVH estimations were obtained by subdividing training set organs-at-risk into high-risk clinical target volume boundary distance subvolumes and computing cohort-averaged differential DVHs. Full DVH estimation was then performed on the training and validation sets. Model performance was quantified by ΔD2cc = D2cc(actual)-D2cc(predicted) (mean and standard deviation). ΔD2cc between training and validation sets were compared with a Student's t test (p < 0.01 significant). Categorical variables (physician, fraction-number, total fractions, and case complexity) that might explain model variance were examined using an analysis of variance test (Bonferroni-corrected p < 0.01 threshold).ResultsTraining set deviations were bladder ΔD2cc = -0.04 ± 0.61 Gy, rectum ΔD2cc = 0.02 ± 0.57 Gy, and sigmoid ΔD2cc = -0.05 ± 0.52 Gy. Model predictions on validation set did not statistically differ: bladder ΔD2cc = -0.02 ± 0.46 Gy (p = 0.80), rectum ΔD2cc = -0.007 ± 0.47 Gy (p = 0.53), and sigmoid ΔD2cc = -0.07 ± 0.47 Gy (p = 0.70). The only significant categorical variable was the attending physician for bladder and rectum ΔD2cc. CONCLUSION: A simple boundary distance-driven knowledge-based DVH estimation exhibited promising results in predicting critical brachytherapy dose metrics. Future work will examine the utility of these predictions for quality control and automated brachytherapy planning.

Objective. To lay the foundation for automated knowledge-based brachytherapy treatment planning using 3D dose estimations, we describe an optimization framework to convert brachytherapy dose distributions directly into dwell times (DTs). Approach. A dose rate kernel d ̇ (r , θ , φ) was produced by exporting 3D dose for one dwell position from the treatment planning system and normalizing by DT. By translating and rotating this kernel to each dwell position, scaling by DT and summing over all dwell positions, dose was computed (D calc). We used a Python-coded COBYLA optimizer to iteratively determine the DTs that minimize the mean squared error between D calc and reference dose D ref, computed using voxels with D ref 80%–120% of prescription. As validation of the optimization, we showed that the optimizer replicates clinical plans when D ref = clinical dose in 40 patients treated with tandem-and-ovoid (T&O) or tandem-and-ring (T&R) and 0–3 needles. Then we demonstrated automated planning in 10 T&O using D ref = dose predicted from a convolutional neural network developed in past work. Validation and automated plans were compared to clinical plans using mean absolute differences ( MAD = 1 N ∑ n = 1 N abs x n − x n ′ ) over all voxels (xn = Dose, N = #voxels) and DTs (xn = DT, N = #dwell positions), mean differences (MD) in organ D 2cc and high-risk CTV D90 over all patients (where positive indicates higher clinical dose), and mean Dice similarity coefficients (DSC) for 100% isodose contours. Main results. Validation plans agreed well with clinical plans (MADdose = 1.1%, MADDT = 4 s or 0.8% of total plan time, D 2cc MD = −0.2% to 0.2% and D90 MD = −0.6%, DSC = 0.99). For automated plans, MADdose = 6.5% and MADDT = 10.3 s (2.1%). The slightly higher clinical metrics in automated plans (D 2cc MD = −3.8% to 1.3% and D90 MD = −5.1%) were due to higher neural network dose predictions. The overall shape of the automated dose distributions were similar to clinical doses (DSC = 0.91). Significance. Automated planning with 3D dose predictions could provide significant time savings and standardize treatment planning across practitioners, regardless of experience. [ABSTRACT FROM AUTHOR]

Purpose: To advance fair and consistent comparisons of dose prediction methods for knowledge‐based planning (KBP) in radiation therapy research. Methods: We hosted OpenKBP, a 2020 AAPM Grand Challenge, and challenged participants to develop the best method for predicting the dose of contoured computed tomography (CT) images. The models were evaluated according to two separate scores: (a) dose score, which evaluates the full three‐dimensional (3D) dose distributions, and (b) dose‐volume histogram (DVH) score, which evaluates a set DVH metrics. We used these scores to quantify the quality of the models based on their out‐of‐sample predictions. To develop and test their models, participants were given the data of 340 patients who were treated for head‐and‐neck cancer with radiation therapy. The data were partitioned into training (n=200), validation (n=40), and testing (n=100) datasets. All participants performed training and validation with the corresponding datasets during the first (validation) phase of the Challenge. In the second (testing) phase, the participants used their model on the testing data to quantify the out‐of‐sample performance, which was hidden from participants and used to determine the final competition ranking. Participants also responded to a survey to summarize their models. Results: The Challenge attracted 195 participants from 28 countries, and 73 of those participants formed 44 teams in the validation phase, which received a total of 1750 submissions. The testing phase garnered submissions from 28 of those teams, which represents 28 unique prediction methods. On average, over the course of the validation phase, participants improved the dose and DVH scores of their models by a factor of 2.7 and 5.7, respectively. In the testing phase one model achieved the best dose score (2.429) and DVH score (1.478), which were both significantly better than the dose score (2.564) and the DVH score (1.529) that was achieved by the runner‐up models. Lastly, many of the top performing teams reported that they used generalizable techniques (e.g., ensembles) to achieve higher performance than their competition. Conclusion: OpenKBP is the first competition for knowledge‐based planning research. The Challenge helped launch the first platform that enables researchers to compare KBP prediction methods fairly and consistently using a large open‐source dataset and standardized metrics. OpenKBP has also democratized KBP research by making it accessible to everyone, which should help accelerate the progress of KBP research. The OpenKBP datasets are available publicly to help benchmark future KBP research. [ABSTRACT FROM AUTHOR]

The adoption of knowledge‐based dose‐volume histogram (DVH) prediction models for assessing organ‐at‐risk (OAR) sparing in radiotherapy necessitates quantification of prediction accuracy and uncertainty. Moreover, DVH prediction error bands should be readily interpretable as confidence intervals in which to find a percentage of clinically acceptable DVHs. In the event such DVH error bands are not available, we present an independent error quantification methodology using a local reference cohort of high‐quality treatment plans, and apply it to two DVH prediction models, ORBIT‐RT and RapidPlan, trained on the same set of 90 volumetric modulated arc therapy (VMAT) plans. Organ‐at‐risk DVH predictions from each model were then generated for a separate set of 45 prostate VMAT plans. Dose‐volume histogram predictions were then compared to their analogous clinical DVHs to define prediction errors Vclin,i‐Vpred,i (ith plan), from which prediction bias μ, prediction error variation σ, and root‐mean‐square error RMSEpred≡1N∑iVclin,i‐Vpred,i2≅σ2+μ2 could be calculated for the cohort. The empirical RMSEpred was then contrasted to the model‐provided DVH error estimates. For all prostate OARs, above 50% Rx dose, ORBIT‐RT μ and σ were comparable to or less than those of RapidPlan. Above 80% Rx dose, μ < 1% and σ < 3‐4% for both models. As a result, above 50% Rx dose, ORBIT‐RT RMSEpred was below that of RapidPlan, indicating slightly improved accuracy in this cohort. Because μ ≈ 0, RMSEpred is readily interpretable as a canonical standard deviation σ, whose error band is expected to correctly predict 68% of normally distributed clinical DVHs. By contrast, RapidPlan's provided error band, although described in literature as a standard deviation range, was slightly less predictive than RMSEpred (55–70% success), while the provided ORBIT‐RT error band was confirmed to resemble an interquartile range (40–65% success) as described. Clinicians can apply this methodology using their own institutions' reference cohorts to (a) independently assess a knowledge‐based model's predictive accuracy of local treatment plans, and (b) interpret from any error band whether further OAR dose sparing is likely attainable. [ABSTRACT FROM AUTHOR]

Purpose: Single-isocenter, volumetric-modulated arc therapy (VMAT) stereotactic radiosurgery (SRS) for multiple brain metastases (multimets) can deliver highly conformal dose distributions and reduce overall patient treatment time compared to other techniques. However, treatment planning for multimet cases is highly complex due to variability in numbers and sizes of brain metastases, as well as their relative proximity to organs-at-risk (OARs). The purpose of this study was to automate the VMAT planning of multimet cases through a knowledge-based planning (KBP) approach that adapts single-target SRS dose predictions to multiple target predictions. Methods: Using a previously published artificial neural network (ANN) KBP system trained on singletarget, linac-based SRS plans, 3D dose distribution predictions for multimet patients were obtained by treating each brain lesion as a solitary target and subsequently combining individual dose predictions into a single distribution. Spatial dose distributions di(...) for each of the i = 1. . .N lesions were merged using the combination function d(...)=[N∑i dni (...)]1=n. The optimal value of n was determined by minimizing root-mean squared (RMS) difference between clinical multimet plans and predicted dose per unit length along the line profile joining each lesion in the clinical cohort. The gradient measure (GM = [3=4π]1=3 [V1=3 50% - V1=3 100%]) is the primary quality metric for SRS plan evaluation at our institution and served as the main comparative metric between clinical plans and the KBP results. A total of 41 previously treated multimet plans, with target numbers ranging from N = 2–10, were used to validate the ANN predictions and subsequent KBP auto-planning routine. Fully deliverable KBP plans were developed by converting predicted dose distribution into patient-specific optimization objectives for the clinical treatment planning system (TPS). Plan parity was maintained through identical arc configuration and target normalization. Overall plan quality improvements were quantified by calculating the difference between SRS quality metrics (QMs): ΔQM = QMclinical − QMKBP. In addition to GM, investigated QMs were: volume of brain receiving ≥ 10 Gy (V10 Gy), volume of brain receiving ≥ 5 Gy (ΔV5 Gy), heterogeneity index (HI), dose to 0.1 cc of the brainstem (D0.1 cc), dose to 1% of the optic chiasm (D1%), and interlesion dose (DIL). In addition to this quantitative analysis, overall plan quality was assessed via blinded plan comparison of the manual and KBP treatment plans by SRS-specializing physicians. Results: A dose combination factor of n = 8 yielded an integrated dose profile RMS difference of 2.9% across the 41-patient cohort. Multimet dose predictions exhibited ΔGM = 0.07 ± 0.10 cm against the clinical sample, implying either further normal tissue sparing was possible or that dose predictions were slightly overestimating achievable dose gradients. The latter is the more likely explanation, as this bias vanished when dose predictions were converted to deliverable KBP plans ΔGM = 0.00 ± 0.08 cm. Remaining QMs were nearly identical or showed modest improvements in the KBP sample. Equivalent QMs included: ΔV10 Gy = 0.37 ± 3.78 cc, ΔHI = 0.02 ± 0.08 and ΔDIL = −2.22 ± 171.4 cGy. The KBP plans showed a greater degree of normal tissue sparing as indicated by brain ΔV5 Gy = 4.11± 24.05 cc, brainstem ΔD0.1 cc = 42.8 ± 121.4 cGy, and chiasm ΔD1% = 50.8 ± 83.0 cGy. In blinded review by SRS-specializing physicians, KBP-generated plans were deemed equivalent or superior in 32/41(78.1%) of the cases. Conclusion: Heuristic KBP-driven automated planning in linac-based, single-isocenter treatments for multiple brain metastases maintained or exceeded overall plan quality. [ABSTRACT FROM AUTHOR]