Incorporating single-side sparing in models for predicting parotid dose sparing in head and neck IMRT

Technical note: Determining the applicability of a clinical knowledge-based learning model via prospective outlier detection

BACKGROUND

Knowledge-based planning (KBP) is increasingly implemented clinically because of its demonstrated ability to improve treatment planning efficiency and reduce plan quality variations. However, cases with large dose-volume histogram (DVH) prediction uncertainties may still need manual adjustments by the planner to achieve high plan quality.

PURPOSE

The purpose of this study is to develop a data-driven method to detect patients with high prediction uncertainties so that intentional effort is directed to these patients.

METHODS

We apply an anomaly detection method known as the local outlier factor (LOF) to a dataset consisting of the training set and each of the prospective patients considered, to evaluate their likelihood of being an anomaly when compared with the training cases. Features used in the LOF analysis include anatomical features and the model-generated DVH principal component scores. To test the efficacy of the proposed model, 142 prostate patients were retrieved from the clinical database and split into a training dataset of 100 patients and a test dataset of 42 patients. The outlier identification performance was quantified by the difference between the DVH prediction root-mean-squared errors (RMSE) of the identified outlier cases and that of the remaining inlier cases.

RESULTS

With a predefined LOF threshold of 1.4, the inlier cases achieved average RMSEs of 5.0 and 6.7 for bladder and rectum, while the outlier cases have substantially higher RMSEs of 6.7 and 13.0 in comparison.

CONCLUSIONS

We propose a method that can determine the prospective patient's outlier status. This method can be integrated into existing automated treatment planning workflows to reduce the risk of generating suboptimal treatment plans while providing an upfront alert to the treatment planner.

Utilizing pre-determined beam orientation information in dose prediction by 3D fully-connected network for intensity modulated radiotherapy

Background

Although the effect of pre-determined beam orientation on dose distribution of intensity modulated radiotherapy (IMRT) has been well-documented, its impacts on dose prediction are less investigated. In this study, the direction map of beam orientation was incorporated into our proposed deep-learning network and utilized in dose prediction of IMRT plans consisting of multiple static fields.

Methods

The direction map was used to characterize the radiation path through region of interest along a beam orientation. Besides, the distance map was used to characterize the spatial distribution between organs at risk (OARs) and planning target volume (PTV). The input of prediction model consisted of CT image, mask image (for PTV and OARs), distance map, and direction map. The output of prediction model was the estimated dose distribution in three dimensions. A 3D fully-connected network composed of a down-sampling encoder and an up-sampling pyramid decoder was trained based on the calculated 3D dose distributions obtained from a treatment planning system. The voxel-level mean absolute error (MAE), dosimetric metrics, and dose-volume histogram were employed to assess the quality of the estimated dose distribution. Performance of the prediction model was evaluated in two aspects. First, the effectiveness of the new features, direction map, distance maps, and pyramid decoder on prediction accuracy of model were assessed. Second, the proposed model was compared with the other three published prediction models, 3D UNet, ResNet-anti-ResNet, U-ResNet-D for inter-model evaluation.

Results

The improvement of prediction accuracy was 0.38 with the input of direction map and 0.43 with the input of distance map. Our proposed model achieved the least MAE (3.97±1.42) compared with the other three models: (5.37±1.51) for ResNet-anti-ResNet, (4.45±1.52) for U-ResNet-D, and (4.53±1.72) for Unet-3D.

Conclusions

The preliminary result demonstrated that the prediction accuracy of the proposed model was higher than those of the other three state-of-the-art prediction models. The introduction of direction maps, distance map, and pyramid decoder can effectively improve the performance of the current deep-learning network-based prediction models.

Knowledge-based radiation treatment planning: A data-driven method survey

This paper surveys the data-driven dose prediction methods investigated for knowledge-based planning (KBP) in the last decade. These methods were classified into two major categories-traditional KBP methods and deep-learning (DL) methods-according to their techniques of utilizing previous knowledge. Traditional KBP methods include studies that require geometric or anatomical features to either find the best-matched case(s) from a repository of prior treatment plans or to build dose prediction models. DL methods include studies that train neural networks to make dose predictions. A comprehensive review of each category is presented, highlighting key features, methods, and their advancements over the years. We separated the cited works according to the framework and cancer site in each category. Finally, we briefly discuss the performance of both traditional KBP methods and DL methods, then discuss future trends of both data-driven KBP methods to dose prediction.

Technical Note: A cascade 3D U-Net for dose prediction in radiotherapy

PURPOSE

Although large datasets are available, to learn a robust dose prediction model from a limited dataset still remains challenging. This work employed cascaded deep learning models and advanced training strategies with a limited dataset to precisely predict three-dimensional (3D) dose distribution.

METHODS

A Cascade 3D (C3D) model is developed based on the cascade mechanism and 3D U-Net network units. During model training, data augmentations are used to improve the generalization ability of the prediction model. A knowledge distillation technique is employed to further improve the capability of model learning. The C3D network was evaluated using the OpenKBP challenge dataset and competed with those models proposed by more than 40 teams globally. Additionally, it was compared with five existing cutting-edge dose prediction models. The performance of these prediction models was evaluated by voxel-based mean absolute error (MAE) and clinical-related dosimetric metrics. The code and models are publicly available online (https://github.com/LSL000UD/RTDosePrediction).

RESULTS

The MAE of a single C3D model without test-time augmentation is 2.50 Gy (3.57% related to prescription dose) for nonzero dose area, which outperforms the other five dose prediction models by about 0.1 Gy-1.7 Gy. The C3D model won both dose and DVH streams of AAPM 2020 OpenKBP challenge with dose score of 2.31 and DVH score of 1.55.

CONCLUSIONS

The Cascading U-Nets is an ideal solution for 3D dose prediction from a limited dataset. The proper data preprocessing, data augmentation, and optimization procedure are more important than architectural modifications of deep learning network.

Predicting voxel-level dose distributions for esophageal radiotherapy using densely connected network with dilated convolutions

This work aims to develop a voxel-level dose prediction framework by integrating distance information between PTV and OARs, as well as image information, into a densely-connected network (DCNN). Firstly, a four-channel feature map, consisting of a PTV image, an OAR image, a CT image, and a distance image, is constructed. A densely connected neural network is then built and trained for voxel-level dose prediction. Considering that the shape and size of OARs are highly inconsistent, a dilated convolution is employed to capture features from multiple scales. Finally, the proposed network is evaluated with five-fold cross-validation, based on ninety-eight clinically approved treatment plans. The voxel-level mean absolute error(MAE _V ) of DCNN was 2.1% for PTV, 4.6% for left lung, 4.0% for right lung, 5.1% for heart, 6.0% for spinal cord, and 3.4% for body, which outperforms conventional U-Net, Resnet-antiResnet, U-Resnet-D by 0.1-0.8%. This result shows that with the introduction of a distance image and DCNN model, the accuracy of predicted dose distribution could be significantly improved. This approach offers a new dose prediction tool to support quality assurance and the automation of treatment planning in esophageal radiotherapy.

A method of using deep learning to predict three-dimensional dose distributions for intensity-modulated radiotherapy of rectal cancer

Purpose

To develop and test a three‐dimensional (3D) deep learning model for predicting 3D voxel‐wise dose distributions for intensity‐modulated radiotherapy (IMRT).

Methods

A total of 122 postoperative rectal cancer cases treated by IMRT were considered in the study, of which 100 cases were randomly selected as the training–validating set and the remaining as the testing set. A 3D deep learning model named 3D U‐Res‐Net_B was constructed to predict 3D dose distributions. Eight types of 3D matrices from CT images, contoured structures, and beam configurations were fed into the independent input channel, respectively, and the 3D matrix of dose distributions was taken as the output to train the 3D model. The obtained 3D model was used to predict new 3D dose distributions. The predicted accuracy was evaluated in two aspects: (a) The dice similarity coefficients (DSCs) of different isodose volumes, the average dose difference of all voxels within the body, and 3%/5 mm global gamma passing rates of organs at risks (OARs) and planned target volume (PTV) were used to address the spatial correspondence between predicted and clinical delivered 3D dose distributions; (b) The dosimetric index (DI) including homogeneity index, conformity index, V₅₀, V₄₅ for PTV and OARs between predicted and clinical truth were statistically analyzed with the paired‐samples t test. The model was also compared with 3D U‐Net and the same architecture model without beam configurations input (named as 3D U‐Res‐Net_O).

Results

The 3D U‐Res‐Net_B model predicted 3D dose distributions accurately. For the 22 testing cases, the average prediction bias ranged from −1.94% to 1.58%, and the overall mean absolute errors (MAEs) was 3.92 ± 4.16%; there was no statistically significant difference for nearly all DIs. The model had a DSCs value above 0.9 for most isodose volumes, and global 3D gamma passing rates varying from 0.81 to 0.90 for PTV and OARs, clearly outperforming 3D U‐Res‐Net_O and being slightly superior to 3D U‐Net.

Conclusions

This study developed a more general deep learning model by considering beam configurations input and achieved an accurate 3D voxel‐wise dose prediction for rectal cancer treated by IMRT, a potentially easier clinical implementation for more comprehensive automatic planning.

Modeling of multiple planning target volumes for head and neck treatments in knowledge-based treatment planning

PURPOSE

The purpose of this study is to develop an accurate and reliable dose volume histogram (DVH) prediction method for external beam radiation therapy plans with multiple planning target volumes (PTVs).

MATERIALS AND METHODS

We present a novel DVH prediction workflow, including new features and a modeling methodology, that makes better use of multiple PTVs: (a) We propose a generalized feature to characterize the geometric relationship of organ-at-risk (OARs) with respect to two or more PTVs with different prescribed dose levels; (b) We incorporate a novel data augmentation method to improve the data distribution in the feature space; (c) A similarity metric that leverages such information is subsequently used to select a subset of similar cases from the training dataset for model building; (d) Finally, a DVH prediction model is trained with these selected cases. To evaluate this new modeling workflow, we used 120 head and neck (HN) cases to tune the model, and used a separate dataset consisting of 148 cases for validation. The proposed model has been compared with the conventional knowledge-based model in terms of model prediction accuracy, which was measured by the root mean squared error (RMSE) between the predicted DVHs and the actual clinical plan DVHs. Furthermore, 25 randomly selected plans were replanned guided by the proposed model and evaluated against clinical plans using clinical evaluation criteria.

RESULTS

The proposed modeling workflow significantly improved DVH prediction accuracy for brainstem (P < 0.001), cord (P < 0.001), larynx (P = 0.004), mandible (P < 0.001), oral cavity (P = 0.011), parotid (P < 0.001) and pharynx (P = 0.001). Cases replanned with the guidance of the proposed model spared OARs significantly better by clinical evaluation criteria. The replanned cases showed a 15% increase in the number of satisfied criteria, compared with clinical plans.

CONCLUSIONS

The proposed modeling workflow generates DVH predictions with improved accuracy and robustness when multiple PTVs exist in a plan. It has demonstrated that the improvement in the DVH prediction model translates into better plan quality in knowledge-based planning.

Knowledge-based planning for intensity-modulated radiation therapy: A review of data-driven approaches

Purpose

Intensity‐Modulated Radiation Therapy (IMRT), including its variations (including IMRT, Volumetric Arc Therapy (VMAT), and Tomotherapy), is a widely used and critically important technology for cancer treatment. It is a knowledge‐intensive technology due not only to its own technical complexity, but also to the inherently conflicting nature of maximizing tumor control while minimizing normal organ damage. As IMRT experience and especially the carefully designed clinical plan data are accumulated during the past two decades, a new set of methods commonly termed knowledge‐based planning (KBP) have been developed that aim to improve the quality and efficiency of IMRT planning by learning from the database of past clinical plans. Some of this development has led to commercial products recently that allowed the investigation of KBP in numerous clinical applications. In this literature review, we will attempt to present a summary of published methods of knowledge‐based approaches in IMRT and recent clinical validation results.

Methods

In March 2018, a literature search was conducted in the NIH Medline database using the PubMed interface to identify publications that describe methods and validations related to KBP in IMRT including variations such as VMAT and Tomotherapy. The search criteria were designed to have a broad scope to capture relevant results with high sensitivity. The authors filtered down the search results according to a predefined selection criteria by reviewing the titles and abstracts first and then by reviewing the full text. A few papers were added to the list based on the references of the reviewed papers. The final set of papers was reviewed and summarized here.

Results

The initial search yielded a total of 740 articles. A careful review of the titles, abstracts, and eventually the full text and then adding relevant articles from reviewing the references resulted in a final list of 73 articles published between 2011 and early 2018. These articles described methods for developing knowledge models for predicting such parameters as dosimetric and dose‐volume points, voxel‐level doses, and objective function weights that improve or automate IMRT planning for various cancer sites, addressing different clinical and quality assurance needs, and using a variety of machine learning approaches. A number of articles reported carefully designed clinical studies that assessed the performance of KBP models in realistic clinical applications. Overwhelming majority of the studies demonstrated the benefits of KBP in achieving comparable and often improved quality of IMRT planning while reducing planning time and plan quality variation.

Conclusions

The number of KBP‐related studies has been steadily increasing since 2011 indicating a growing interest in applying this approach to clinical applications. Validation studies have generally shown KBP to produce plans with quality comparable to expert planners while reducing the time and efforts to generate plans. However, current studies are mostly retrospective and leverage relatively small datasets. Larger datasets collected through multi‐institutional collaboration will enable the development of more advanced models to further improve the performance of KBP in complex clinical cases. Prospective studies will be an important next step toward widespread adoption of this exciting technology.

Stochastic frontier analysis as knowledge-based model to improve sparing of organs-at-risk for VMAT-treated prostate cancer

Stochastic frontier analysis (SFA) is used as a novel knowledge-based technique in order to develop a predictive model of dosimetric features from significant geometric parameters describing a patient morphology. 406 patients treated with VMAT for prostate cancer were analyzed retrospectively. Cases were divided into three prescription-based groups. Seven geometric parameters are extracted to characterize the relationship between the organs-at-risk (bladder and rectum) with the planning volume (PTV). In total, 37 dosimetric parameters are tested for these two OARs. SFA allows the determination of the minimum achievable dose to the OAR based on the geometric parameters. Stochastic frontiers are determined with a maximum likelihood estimation technique. The SFA model was tested using validation cohort (30 patients with prescribed dose between 60 and 70 Gy) where 77% (23 out of 30) of the predicted DVHs present a 5% or less dose deterioration for the bladder and rectum with the planned DVH. SFA can be used in EBRT planning as a predictive model based on anatomical features of previously treated plans.

Incorporating Case-Based Reasoning for Radiation Therapy Knowledge Modeling: A Pelvic Case Study

Knowledge models in radiotherapy capture the relation between patient anatomy and dosimetry to provide treatment planning guidance. When treatment schemes evolve, existing models struggle to predict accurately. We propose a case-based reasoning framework designed to handle novel anatomies that are of same type but vary beyond original training samples. A total of 105 pelvic intensity-modulated radiotherapy cases were analyzed. Eighty cases were prostate cases while the other 25 were prostate-plus-lymph-node cases. We simulated 4 scenarios: Scarce scenario, Semiscarce scenario, Semiample scenario, and Ample scenario. For the Scarce scenario, a multiple stepwise regression model was trained using 85 cases (80 prostate, 5 prostate-plus-lymph-node). The proposed workflow started with evaluating the feature novelty of new cases against 5 training prostate-plus-lymph-node cases using leverage statistic. The case database was composed of a 5-case dose atlas. Case-based dose prediction was compared against the regression model prediction using sum of squared residual. Mean sum of squared residual of case-based and regression predictions for the bladder of 13 identified outliers were 0.174 ± 0.166 and 0.459 ± 0.508, respectively (P = .0326). For the rectum, the respective mean sum of squared residuals were 0.103 ± 0.120 and 0.150 ± 0.171 for case-based and regression prediction (P = .1972). By retaining novel cases, under the Ample scenario, significant statistical improvement was observed over the Scarce scenario (P = .0398) for the bladder model. We expect that the incorporation of case-based reasoning that judiciously applies appropriate predictive models could improve overall prediction accuracy and robustness in clinical practice.

Segmentation of parotid glands from registered CT and MR images

PURPOSE

To develop an automatic multimodal method for segmentation of parotid glands (PGs) from pre-registered computed tomography (CT) and magnetic resonance (MR) images and compare its results to the results of an existing state-of-the-art algorithm that segments PGs from CT images only.

METHODS

Magnetic resonance images of head and neck were registered to the accompanying CT images using two different state-of-the-art registration procedures. The reference domains of registered image pairs were divided on the complementary PG regions and backgrounds according to the manual delineation of PGs on CT images, provided by a physician. Patches of intensity values from both image modalities, centered around randomly sampled voxels from the reference domain, served as positive or negative samples in the training of the convolutional neural network (CNN) classifier. The trained CNN accepted a previously unseen (registered) image pair and classified its voxels according to the resemblance of its patches to the patches used for training. The final segmentation was refined using a graph-cut algorithm, followed by the dilate-erode operations.

RESULTS

Using the same image dataset, segmentation of PGs was performed using the proposed multimodal algorithm and an existing monomodal algorithm, which segments PGs from CT images only. The mean value of the achieved Dice overlapping coefficient for the proposed algorithm was 78.8%, while the corresponding mean value for the monomodal algorithm was 76.5%.

CONCLUSIONS

Automatic PG segmentation on the planning CT image can be augmented with the MR image modality, leading to an improved RT planning of head and neck cancer.

Creation of knowledge-based planning models intended for large scale distribution: Minimizing the effect of outlier plans

Knowledge‐based planning (KBP) can be used to estimate dose–volume histograms (DVHs) of organs at risk (OAR) using models. The task of model creation, however, can result in estimates with differing accuracy; particularly when outlier plans are not properly addressed. This work used RapidPlan^™ to create models for the prostate and head and neck intended for large‐scale distribution. Potential outlier plans were identified by means of regression analysis scatter plots, Cook's distance, coefficient of determination, and the chi‐squared test. Outlier plans were identified as falling into three categories: geometric, dosimetric, and over‐fitting outliers. The models were validated by comparing DVHs estimated by the model with those from a separate and independent set of clinical plans. The estimated DVHs were also used as optimization objectives during inverse planning. The analysis tools lead us to identify as many as 7 geometric, 8 dosimetric, and 20 over‐fitting outliers in the raw models. Geometric and over‐fitting outliers were removed while the dosimetric outliers were replaced after re‐planning. Model validation was done by comparing the DVHs at 50%, 85%, and 99% of the maximum dose for each OAR (denoted as V50, V85, and V99) and agreed within −2% to 4% for the three metrics for the final prostate model. In terms of the head and neck model, the estimated DVHs agreed from −2.0% to 5.1% at V50, 0.1% to 7.1% at V85, and 0.1% to 7.6% at V99. The process used to create these models improved the accuracy for the pharyngeal constrictor DVH estimation where one plan was originally over‐estimated by more than twice. In conclusion, our results demonstrate that KBP models should be carefully created since their accuracy could be negatively affected by outlier plans. Outlier plans can be addressed by removing them from the model and re‐planning.

An Ensemble Approach to Knowledge-Based Intensity-Modulated Radiation Therapy Planning

Knowledge-based planning (KBP) utilizes experienced planners’ knowledge embedded in prior plans to estimate optimal achievable dose volume histogram (DVH) of new cases. In the regression-based KBP framework, previously planned patients’ anatomical features and DVHs are extracted, and prior knowledge is summarized as the regression coefficients that transform features to organ-at-risk DVH predictions. In our study, we find that in different settings, different regression methods work better. To improve the robustness of KBP models, we propose an ensemble method that combines the strengths of various linear regression models, including stepwise, lasso, elastic net, and ridge regression. In the ensemble approach, we first obtain individual model prediction metadata using in-training-set leave-one-out cross validation. A constrained optimization is subsequently performed to decide individual model weights. The metadata is also used to filter out impactful training set outliers. We evaluate our method on a fresh set of retrospectively retrieved anonymized prostate intensity-modulated radiation therapy (IMRT) cases and head and neck IMRT cases. The proposed approach is more robust against small training set size, wrongly labeled cases, and dosimetric inferior plans, compared with other individual models. In summary, we believe the improved robustness makes the proposed method more suitable for clinical settings than individual models.

A method for a priori estimation of best feasible DVH for organs-at-risk: Validation for head and neck VMAT planning

PURPOSE

Despite improvements in optimization and automation algorithms, the quality of radiation treatment plans still varies dramatically. A tool that allows a priori estimation of the best possible sparing (Feasibility DVH, or FDVH) of an organ at risk (OAR) in high-energy photon planning may help reduce plan quality variability by deriving patient-specific OAR goals prior to optimization. Such a tool may be useful for (a) meaningfully evaluating patient-specific plan quality and (b) supplying best theoretically achievable DVH goals, thus pushing the solution toward automatic Pareto optimality. This work introduces such a tool and validates it for clinical Head and Neck (HN) datasets.

METHODS

To compute FDVH, first the targets are assigned uniform prescription doses, with no reference to any particular beam arrangement. A benchmark 3D dose built outside the targets is estimated using a series of energy-specific dose spread calculations reflecting observed properties of radiation distribution in media. For the patient, the calculation is performed on the heterogeneous dataset, taking into account the high- (penumbra driven) and low- (PDD and scatter-driven) gradient dose spreading. The former is driven mostly by target dose and surface shape, while the latter adds the dependence on target volume. This benchmark dose is used to produce the "best possible sparing" FDVH for an OAR, and based on it, progressively more easily achievable FDVH curves can be estimated. Validation was performed using test cylindrical geometries as well as 10 clinical HN datasets. For HN, VMAT plans were prepared with objectives of covering the primary and the secondary (bilateral elective neck) PTVs while addressing only one OAR at a time, with the goal of maximum sparing. The OARs were each parotid, the larynx, and the inferior pharyngeal constrictor. The difference in mean OAR doses was computed for the achieved vs. FDVHs, and the shapes of those DVHs were compared by means of the Dice similarity coefficient (DSC).

RESULTS

For all individually optimized HN OARs (N = 38), the average DSC between the planned DVHs and the FDVHs was 0.961 ± 0.018 (95% CI 0.955-0.967), with the corresponding average of mean OAR dose differences of 1.8 ± 5.8% (CI -0.1-3.6%). For realistic plans the achieved DVHs run no lower than the FDVHs, except when target coverage is compromised at the target/OAR interface.

CONCLUSIONS

For the validation of VMAT plans, the OAR DVHs optimized one-at-a-time were similar in shape to and bound on the low side by the FDVHs, within the confines of planner's ability to precisely cover the target(s) with the prescription dose(s). The method is best suited for the OARs close to the target. This approach is fundamentally different from "knowledge-based planning" because it is (a) independent of the treatment plan and prior experience, and (b) it approximates, from nearly first principles, the lowest possible boundary of the OAR DVH, but not necessarily its actual shape in the presence of competing OAR sparing and target dose homogeneity objectives.

Highly Efficient Training, Refinement, and Validation of a Knowledge-based Planning Quality-Control System for Radiation Therapy Clinical Trials

PURPOSE

To demonstrate an efficient method for training and validation of a knowledge-based planning (KBP) system as a radiation therapy clinical trial plan quality-control system.

METHODS AND MATERIALS

We analyzed 86 patients with stage IB through IVA cervical cancer treated with intensity modulated radiation therapy at 2 institutions according to the standards of the INTERTECC (International Evaluation of Radiotherapy Technology Effectiveness in Cervical Cancer, National Clinical Trials Network identifier: 01554397) protocol. The protocol used a planning target volume and 2 primary organs at risk: pelvic bone marrow (PBM) and bowel. Secondary organs at risk were rectum and bladder. Initial unfiltered dose-volume histogram (DVH) estimation models were trained using all 86 plans. Refined training sets were created by removing sub-optimal plans from the unfiltered sample, and DVH estimation models… and DVH estimation models were constructed by identifying 30 of 86 plans emphasizing PBM sparing (comparing protocol-specified dosimetric cutpoints V₁₀ (percentage volume of PBM receiving at least 10 Gy dose) and V₂₀ (percentage volume of PBM receiving at least 20 Gy dose) with unfiltered predictions) and another 30 of 86 plans emphasizing bowel sparing (comparing V₄₀ (absolute volume of bowel receiving at least 40 Gy dose) and V₄₅ (absolute volume of bowel receiving at least 45 Gy dose), 9 in common with the PBM set). To obtain deliverable KBP plans, refined models must inform patient-specific optimization objectives and/or priorities (an auto-planning "routine"). Four candidate routines emphasizing different tradeoffs were composed, and a script was developed to automatically re-plan multiple patients with each routine. After selection of the routine that best met protocol objectives in the 51-patient training sample (KBP_FINAL), protocol-specific DVH metrics and normal tissue complication probability were compared for original versus KBP_FINAL plans across the 35-patient validation set. Paired t tests were used to test differences between planning sets.

RESULTS

KBP_FINAL plans outperformed manual planning across the validation set in all protocol-specific DVH cutpoints. The mean normal tissue complication probability for gastrointestinal toxicity was lower for KBP_FINAL versus validation-set plans (48.7% vs 53.8%, P<.001). Similarly, the estimated mean white blood cell count nadir was higher (2.77 vs 2.49 k/mL, P<.001) with KBP_FINAL plans, indicating lowered probability of hematologic toxicity.

CONCLUSIONS

This work demonstrates that a KBP system can be efficiently trained and refined for use in radiation therapy clinical trials with minimal effort. This patient-specific plan quality control resulted in improvements on protocol-specific dosimetric endpoints.

From active shape model to active optical flow model: a shape-based approach to predicting voxel-level dose distributions in spine SBRT

Prediction of achievable dose distribution in spine stereotactic body radiation therapy (SBRT) can help in designing high-quality treatment plans to maximally protect spinal cords and to effectively control tumours. Dose distributions at spinal cords are primarily affected by the shapes of adjacent planning target volume (PTV) contours. In this work, we estimate such contour effects and predict dose distributions by exploring active optical flow model (AOFM) and active shape model (ASM). We first collect a sequence of dose sub-images and PTV contours near spinal cords from fifteen SBRT plans in the training dataset. The data collection is then classified into five groups according to the PTV locations in relation to spinal cords. In each group, we randomly choose a dose sub-image as the reference and register all other sub-images to the reference using an optical flow method. AOFM is then constructed by importing optical flow vectors and dose values into the principal component analysis (PCA). Similarly, we build ASM by using PCA on PTV contour points. The correlation between ASM and AOFM is estimated via a stepwise multiple regression model. When predicting dose distribution of a new case, the group is first determined based on the PTV contour. The prediction model of the selected group is used to estimate dose distributions by mapping the PTV contours from the ASM space to the AOFM space. This method was validated on fifteen SBRT plans in the testing dataset. Analysis of dose-volume histograms revealed that the important D2%, D5%, D10% and D0.1cc dosimetric parameters of spinal cords between the prediction and the clinical plans were 11.7 ± 1.7 Gy versus 11.8 ± 1.7 Gy (p = 0.95), 10.9 ± 1.7 Gy versus 11.1 ± 1.9 Gy (p = 0.8295), 10.2 ± 1.6 Gy versus 10.1 ± 1.7 (p = 0.9036) and 11.2 ± 2.0 Gy versus 11.1 ± 2.2 Gy (p = 0.5208), respectively. Here, the ‘cord’ is the spinal cord proper (not the thecal sac) extended 5 mm inferior and superior to the involved vertebral bodies, and the ‘PTV’ is the involved segment of the vertebral body expanded uniformly by 2 mm but excluding the spinal cord volume expanded by 2 mm (Ref. RTOG 0631). These results suggested that the AOFM-based approach is a promising tool for predicting accurate spinal cord dose in clinical practice. In this work, we demonstrated the feasibility of using AOFM and ASM models derived from previously treated patients to estimate the achievable dose distributions for new patients.

Knowledge-based prediction of plan quality metrics in intracranial stereotactic radiosurgery

PURPOSE

The objective of this work was to develop a comprehensive knowledge-based methodology for predicting achievable dose-volume histograms (DVHs) and highly precise DVH-based quality metrics (QMs) in stereotactic radiosurgery/radiotherapy (SRS/SRT) plans. Accurate QM estimation can identify suboptimal treatment plans and provide target optimization objectives to standardize and improve treatment planning.

METHODS

Correlating observed dose as it relates to the geometric relationship of organs-at-risk (OARs) to planning target volumes (PTVs) yields mathematical models to predict achievable DVHs. In SRS, DVH-based QMs such as brain V10Gy (volume receiving 10 Gy or more), gradient measure (GM), and conformity index (CI) are used to evaluate plan quality. This study encompasses 223 linear accelerator-based SRS/SRT treatment plans (SRS plans) using volumetric-modulated arc therapy (VMAT), representing 95% of the institution's VMAT radiosurgery load from the past four and a half years. Unfiltered models that use all available plans for the model training were built for each category with a stratification scheme based on target and OAR characteristics determined emergently through initial modeling process. Model predictive accuracy is measured by the mean and standard deviation of the difference between clinical and predicted QMs, δQM = QMclin - QMpred, and a coefficient of determination, R(2). For categories with a large number of plans, refined models are constructed by automatic elimination of suspected suboptimal plans from the training set. Using the refined model as a presumed achievable standard, potentially suboptimal plans are identified. Predictions of QM improvement are validated via standardized replanning of 20 suspected suboptimal plans based on dosimetric predictions. The significance of the QM improvement is evaluated using the Wilcoxon signed rank test.

RESULTS

The most accurate predictions are obtained when plans are stratified based on proximity to OARs and their PTV volume sizes. Volumes are categorized into small (VPTV < 2 cm(3)), medium (2 cm(3) < VPTV < 25 cm(3)), and large (25 cm(3) < VPTV). The unfiltered models demonstrate the ability to predict GMs to ∼1 mm and fractional brain V10Gy to ∼25% for plans with large VPTV and critical OAR involvements. Increased accuracy and precision of QM predictions are obtained when high quality plans are selected for the model training. For the small and medium VPTV plans without critical OAR involvement, predictive ability was evaluated using the refined model. For training plans, the model predicted GM to an accuracy of 0.2 ± 0.3 mm and fractional brain V10Gy to 0.04 ± 0.12, suggesting highly accurate predictive ability. For excluded plans, the average δGM was 1.1 mm and fractional brain V10Gy was 0.20. These δQM are significantly greater than those of the model training plans (p < 0.001). For CI, predictions are close to clinical values and no significant difference was observed between the training and excluded plans (p = 0.19). Twenty outliers with δGM > 1.35 mm were identified as potentially suboptimal, and replanning these cases using predicted target objectives demonstrates significant improvements on QMs: on average, 1.1 mm reduction in GM (p < 0.001) and 23% reduction in brain V10Gy (p < 0.001). After replanning, the difference of δGM distribution between the 20 replans and the refined model training plans was marginal.

CONCLUSIONS

The results demonstrate the ability to predict SRS QMs precisely and to identify suboptimal plans. Furthermore, the knowledge-based DVH predictions were directly used as target optimization objectives and allowed a standardized planning process that bettered the clinically approved plans. Full clinical application of this methodology can improve consistency of SRS plan quality in a wide range of PTV volume and proximity to OARs and facilitate automated treatment planning for this critical treatment site.

Patient feature based dosimetric Pareto front prediction in esophageal cancer radiotherapy

PURPOSE

To investigate the feasibility of the dosimetric Pareto front (PF) prediction based on patient's anatomic and dosimetric parameters for esophageal cancer patients.

METHODS

Eighty esophagus patients in the authors' institution were enrolled in this study. A total of 2928 intensity-modulated radiotherapy plans were obtained and used to generate PF for each patient. On average, each patient had 36.6 plans. The anatomic and dosimetric features were extracted from these plans. The mean lung dose (MLD), mean heart dose (MHD), spinal cord max dose, and PTV homogeneity index were recorded for each plan. Principal component analysis was used to extract overlap volume histogram (OVH) features between PTV and other organs at risk. The full dataset was separated into two parts; a training dataset and a validation dataset. The prediction outcomes were the MHD and MLD. The spearman's rank correlation coefficient was used to evaluate the correlation between the anatomical features and dosimetric features. The stepwise multiple regression method was used to fit the PF. The cross validation method was used to evaluate the model.

RESULTS

With 1000 repetitions, the mean prediction error of the MHD was 469 cGy. The most correlated factor was the first principal components of the OVH between heart and PTV and the overlap between heart and PTV in Z-axis. The mean prediction error of the MLD was 284 cGy. The most correlated factors were the first principal components of the OVH between heart and PTV and the overlap between lung and PTV in Z-axis.

CONCLUSIONS

It is feasible to use patients' anatomic and dosimetric features to generate a predicted Pareto front. Additional samples and further studies are required improve the prediction model.