Machine learning models to predict the delivered positions of Elekta multileaf collimator leaves for volumetric modulated arc therapy.

PURPOSE: Accurate positioning of multileaf collimator (MLC) leaves during volumetric modulated arc therapy (VMAT) is essential for accurate treatment delivery. We developed a linear regression, support vector machine, random forest, extreme gradient boosting (XGBoost), and an artificial neural network (ANN) for predicting the delivered leaf positions for VMAT plans. METHODS: For this study, 160 MLC log files from 80 VMAT plans were obtained from a single institution treated on 3 Elekta Versa HD linear accelerators. The gravity vector, X1 and X2 jaw positions, leaf gap, leaf position, leaf velocity, and leaf acceleration were extracted and used as model inputs. The models were trained using 70% of the log files and tested on the remaining 30%. Mean absolute error (MAE), root mean square error (RMSE), the coefficient of determination R2 , and fitted line plots showing the relationship between delivered and predicted leaf positions were used to evaluate model performance. RESULTS: The models achieved the following errors: linear regression (MAE = 0.158 mm, RMSE = 0.225 mm), support vector machine (MAE = 0.141 mm, RMSE = 0.199 mm), random forest (MAE = 0.161 mm, RMSE = 0.229 mm), XGBoost (MAE = 0.185 mm, RMSE = 0.273 mm), and ANN (MAE = 0.361 mm, RMSE = 0.521 mm). A significant correlation between a plan's gamma passing rate (GPR) and the prediction errors of linear regression, support vector machine, and random forest is seen (p < 0.045). CONCLUSIONS: We examined various models to predict the delivered MLC positions for VMAT plans treated with Elekta linacs. Linear regression, support vector machine, random forest, and XGBoost achieved lower errors than ANN. Models that can accurately predict the individual leaf positions during treatment can help identify leaves that are deviating from the planned position, which can improve a plan's GPR.

Impacts of COVID-19 social distancing policies on water demand: A population dynamics perspective.

Social distancing policies (SDPs) implemented in response to the COVID-19 pandemic have led to temporal and spatial shifts in water demand across cities. Water utilities need to understand these demand shifts to respond to potential operational and water-quality issues. Aided by a fixed-effects model of citywide water demand in Austin, Texas, we explore the impacts of various SDPs (e.g., time after the stay home-work safe order, reopening phases) using daily demand data gathered between 2013 and 2020. Our approach uses socio-technical determinants (e.g., climate, water conservation policy) with SDPs to model water demand, while accounting for spatial and temporal effects (e.g., geographic variations, weekday patterns). Results indicate shifts in behavior of residential and nonresidential demands that offset the change at the system scale, demonstrating a spatial redistribution of water demand after the stay home-work safe order. Our results show that some phases of Texas's reopening phases had statistically significant relationships to water demand. While this yielded only marginal net effects on overall demand, it underscores behavioral changes in demand at sub-system spatial scales. Our discussions shed light on SDPs' impacts on water demand. Equipped with our empirical findings, utilities can respond to potential vulnerabilities in their systems, such as water-quality problems that may be related to changes in water pressure in response to demand variations.

Leveraging water-wastewater data interdependencies to understand infrastructure systems' behaviors during COVID-19 pandemic.

Social distancing policies (SDPs) implemented worldwide in response to COVID-19 pandemic have led to spatiotemporal variations in water demand and wastewater flow, creating potential operational and service-related quality issues in water-sector infrastructure. Understanding water-demand variations is especially challenging in contexts with limited availability of smart meter infrastructure, hindering utilities' ability to respond in real time to identified system vulnerabilities. Leveraging water and wastewater infrastructures' interdependencies, this study proposes the use of high-granular wastewater-flow data as a proxy to understand both water and wastewater systems' behaviors during active SDPs. Enabled by a random-effects model of wastewater flow in an urban metropolitan city in Texas, we explore the impacts of various SDPs (e.g., stay home-work safe, reopening phases) using daily flow data gathered between March 19, 2019, and December 31, 2020. Results indicate an increase in residential flow that offset a decrease in nonresidential flow, demonstrating a spatial redistribution of wastewater flow during the stay home-work safe period. Our results show that the three reopening phases had statistically significant relationships to wastewater flow. While this yielded only marginal net effects on overall wastewater flow, it serves as an indicator of behavioral changes in water demand at sub-system spatial scales given demand-flow interdependencies. Our assessment should enable utilities without smart meters in their water system to proactively target their operational response during pandemics, such as (1) monitoring wastewater-flow velocity to alleviate potential blockages in sewer pipes in case of decreased flows, and (2) closely investigating any consequential water-quality problems due to decreased demands.

DVH Analytics: A DVH database for clinicians and researchers.

In this study, we build a vendor-agnostic software application capable of importing and analyzing non-image-based DICOM files for various radiation treatment modalities (i.e., DICOM RT Dose, RT Structure, and RT Plan files). Dose-volume histogram (DVH) and planning data are imported into a SQL database, and methods are provided to manage, edit, view, and download data. Furthermore, the software provides various analytical tools for plan evaluations, plan comparisons, benchmarking, and plan outcome predictions. DVH Analytics is developed using Python, including libraries such as pydicom, dicompyler, psycopg2, SciPy, Statsmodels, and Bokeh for parsing DICOM files, computing DVHs, communicating with a PostgreSQL database, performing statistical analyses, and creating a web-based user interface. This software is open-source and compatible with Windows, Mac OS, and Linux. For proof-of-concept, a database with over 3,000 DVHs from a single physician's head & neck practice was built. From these data, differences in means, correlations, and temporal trends in dose to multiple organs-at-risk (OARs) were observed. Furthermore, an example of the predictive regression tool is reported, where a model was constructed to predict maximum dose to brainstem based on minimum distance from planning target volume (PTV) and treatment beam source-to-skin distance (SSD). With DVH Analytics, we have developed a free, open-source software program to parse, organize, and analyze non-image-based DICOM data for use in a radiation oncology setting. Furthermore, this software can be used to generate statistical models for the purposes of quality control or outcome predictions and correlations.

On correlations in IMRT planning aims.

The purpose was to study correlations amongst IMRT DVH evaluation points and how their relaxation impacts the overall plan. 100 head-and-neck cancer cases, using the Eclipse treatment planning system with the same protocol, are statisti-cally analyzed for PTV, brainstem, and spinal cord. To measure variations amongst the plans, we use (i) interquartile range (IQR) of volume as a function of dose, (ii) interquartile range of dose as a function of volume, and (iii) dose falloff. To determine correlations for institutional and ICRU goals, conditional probabilities and medians are computed. We observe that most plans exceed the median PTV dose (average D50 = 104% prescribed dose). Furthermore, satisfying D50 reduced the probability of also satisfying D98, constituting a negative correlation of these goals. On the other hand, satisfying D50 increased the probability of satisfying D2, suggesting a positive correlation. A positive correlation is also observed between the PTV V105 and V110. Similarly, a positive correlation between the brainstem V45 and V50 is measured by an increase in the conditional median of V45, when V50 is violated. Despite the imposed institutional and international recommenda-tions, significant variations amongst DVH points can occur. Even though DVH aims are evaluated independently, sizable correlations amongst them are possible, indicating that some goals cannot be satisfied concurrently, calling for unbiased plan criteria.

Artificial intelligence and machine learning: Definition of terms and current concepts in critical care research.

With increasing computing power, artificial intelligence (AI) and machine learning (ML) have prospered, which facilitate the analysis of large datasets, especially those found in critical care. It is important to define these terminologies, to inform a standardized approach to critical care research. This manuscript hopes to clarify these terms with examples from medical literature. Three major components that are required for a successful ML implementation: (i) reliable dataset, (ii) ML algorithm, and (iii) unbiased model evaluation, are discussed. A reliable dataset can be structured or unstructured with limited noise, outliers, and missing values. ML, a subset of AI, is typically focused on supervised or unsupervised learning tasks in which the output is based on inputs and derived from iterative pattern recognition algorithms, while AI is the overall ability of a machine to "think" or mimic human behavior; and to analyze data free from human influence. Even with successful implementation, advanced AI and ML algorithms have faced challenges in adoption into practice, mainly due to their lack of interpretability, which hinders trust, buy-in, and engagement from clinicians. Consequently, traditional algorithms, such as linear and logistic regression, that may have reduced predictive power but are highly interpretable, continue to be widely used.

Phase I quality control framework for monitoring organ-at-risk dose.

Objective.The aim of this work was to develop a Phase I control chart framework for the recently proposed multivariate risk-adjusted Hotelling'sT2chart. Although this control chart alone can identify most patients receiving extreme organ-at-risk (OAR) dose, it is restricted by underlying distributional assumptions, making it sensitive to extreme observations in the sample, as is typically found in radiotherapy plan quality data such as dose-volume histogram (DVH) points. This can lead to slightly poor-quality plans that should have been identified as out-of-control (OC) to be signaled in-control (IC).Approach. We develop a robust iterative control chart framework to identify all OC patients with abnormally high OAR dose and improve them via re-optimization to achieve an IC sample prior to establishing the Phase I control chart, which can be used to monitor future treatment plans.Main Results. Eighty head-and-neck patients were used in this study. After the first iteration, P14, P67, and P68 were detected as OC for high brainstem dose, warranting re-optimization aimed to reduce brainstem dose without worsening other planning criteria. The DVH and control chart were updated after re-optimization. On the second iteration, P14, P67, and P68 were IC, but P40 was identified as OC. After re-optimizing P40's plan and updating the DVH and control chart, P40 was IC, but P14* (P14's re-optimized plan) and P62 were flagged as OC. P14* could not be re-optimized without worsening target coverage, so only P62 was re-optimized. Ultimately, a fully IC sample was achieved. Multiple iterations were needed to identify and improve all OC patients, and to establish a more robust control limit to monitor future treatment plans.Significance. The iterative procedure resulted in a fully IC sample of patients. With this sample, a more robust Phase I control chart that can monitor OAR doses of new plans was established.

CBCT-guided adaptive radiotherapy using self-supervised sequential domain adaptation with uncertainty estimation.

Adaptive radiotherapy (ART) is an advanced technology in modern cancer treatment that incorporates progressive changes in patient anatomy into active plan/dose adaption during the fractionated treatment. However, the clinical application relies on the accurate segmentation of cancer tumors on low-quality on-board images, which has posed challenges for both manual delineation and deep learning-based models. In this paper, we propose a novel sequence transduction deep neural network with an attention mechanism to learn the shrinkage of the cancer tumor based on patients' weekly cone-beam computed tomography (CBCT). We design a self-supervised domain adaption (SDA) method to learn and adapt the rich textural and spatial features from pre-treatment high-quality computed tomography (CT) to CBCT modality in order to address the poor image quality and lack of labels. We also provide uncertainty estimation for sequential segmentation, which aids not only in the risk management of treatment planning but also in the calibration and reliability of the model. Our experimental results based on a clinical non-small cell lung cancer (NSCLC) dataset with sixteen patients and ninety-six longitudinal CBCTs show that our model correctly learns weekly deformation of the tumor over time with an average dice score of 0.92 on the immediate next step, and is able to predict multiple steps (up to 5 weeks) for future patient treatments with an average dice score reduction of 0.05. By incorporating the tumor shrinkage predictions into a weekly re-planning strategy, our proposed method demonstrates a significant decrease in the risk of radiation-induced pneumonitis up to 35% while maintaining the high tumor control probability.

Managing tumor changes during radiotherapy using a deep learning model.

PURPOSE: We propose a treatment planning framework that accounts for weekly lung tumor shrinkage using cone beam computed tomography (CBCT) images with a deep learning-based model. METHODS: Sixteen patients with non-small-cell lung cancer (NSCLC) were selected with one planning CT and six weekly CBCTs each. A deep learning-based model was applied to predict the weekly deformation of the primary tumor based on the spatial and temporal features extracted from previous weekly CBCTs. Starting from Week 3, the tumor contour at Week N was predicted by the model based on the input from all the previous weeks (1, 2 … N - 1), and was evaluated against the manually contoured tumor using Dice coefficient (DSC), precision, average surface distance (ASD), and Hausdorff distance (HD). Information about the predicted tumor was then entered into the treatment planning system and the plan was re-optimized every week. The objectives were to maximize the dose coverage in the target region while minimizing the toxicity to the surrounding healthy tissue. Dosimetric evaluation of the target and organs at risk (heart, lung, esophagus, and spinal cord) was performed on four cases, comparing between a conventional plan (ignoring tumor shrinkage) and the shrinkage-based plan. RESULTS: he primary tumor volumes decreased on average by 38% ± 26% during six weeks of treatment. DSCs and ASD between the predicted tumor and the actual tumor for Weeks 3, 4, 5, 6 were 0.81, 0.82, 0.79, 0.78 and 1.49, 1.59, 1.92, 2.12 mm, respectively, which were significantly superior to the score of 0.70, 0.68, 0.66, 0.63 and 2.81, 3.22, 3.69, 3.63 mm between the rigidly transferred tumors ignoring shrinkage and the actual tumor. While target coverage metrics were maintained for the re-optimized plans, lung mean dose dropped by 2.85, 0.46, 2.39, and 1.48 Gy for four sample cases when compared to the original plan. Doses in other organs such as esophagus were also reduced for some cases. CONCLUSION: We developed a deep learning-based model for tumor shrinkage prediction. This model used CBCTs and contours from previous weeks as input and produced reasonable tumor contours with a high prediction accuracy (DSC, precision, HD, and ASD). The proposed framework maintained target coverage while reducing dose in the lungs and esophagus.

Treatment plan quality control using multivariate control charts.

PURPOSE: Statistical process control tools such as control charts were recommended by the American Association of Physicists in Medicine (AAPM) Task Group 218 for radiotherapy quality assurance. However, the tools needed to analyze multivariate, correlated data that are often encountered in treatment plan quality measures, are lacking. In this study, we develop quality control tools that can model multivariate plan quality measures with correlations and account for patient-specific risk factors, without adding a significant burden to clinical workflow. METHODS AND MATERIALS: A multivariate, quality control chart is developed that includes a risk-adjustment model, Hotelling's T2 statistic, and principal component analysis (PCA). Principal component analysis accounts for correlations among a set of organ-at-risk (OAR) dose-volume histogram (DVH) points that serves as proxies for plan quality. Risk-adjustment models estimate the principal components from PCA using a set of patient- and treatment-specific risk factors. The resulting residuals from the risk-adjustment models are used to compute the Hotelling's T2 statistic; the corresponding multivariate control chart is then plotted based on the beta distribution followed by the statistic. Further, the box-cox transformation is used to account for non-normality in DVH points. We investigate the application of the proposed methodology via three multivariate control charts - a conventional chart that ignores risk-adjustment and PCA, a risk-adjusted chart ignoring PCA, and a PCA-based, risk-adjusted chart. These control charts are evaluated on 69 head-and-neck cases. RESULTS: The conventional multivariate control chart fails to account for important patient-specific risk factors, including volumes and cross-sectional areas of the tumor and OARs and distances in-between. This failure leads to a larger number of false alarms. While the multivariate risk-adjusted control chart is able to reduce false alarms, it fails to account for correlations in DVH points. The multivariate PCA-based, risk-adjusted control chart can detect unusual plans after accounting for the correlations. By replanning, improvements are shown on an unusual plan identified by both risk-adjusted methods. CONCLUSIONS: The multivariate risk-adjusted control chart developed here enables quality control of plans prior to delivery. This methodology is generic and can be readily applied for other radiotherapy quality assurance protocols, such as gamma analysis pass rates.

A Risk-Adjusted Control Chart to Evaluate Intensity Modulated Radiation Therapy Plan Quality.

PURPOSE: This study aimed to develop a quality control framework for intensity modulated radiation therapy plan evaluations that can account for variations in patient- and treatment-specific risk factors. METHODS AND MATERIALS: Patient-specific risk factors, such as a patient's anatomy and tumor dose requirements, affect organs-at-risk (OARs) dose-volume histograms (DVHs), which in turn affects plan quality and can potentially cause adverse effects. Treatment-specific risk factors, such as the use of chemotherapy and surgery, are clinically relevant when evaluating radiation therapy planning criteria. A risk-adjusted control chart was developed to identify unusual plan quality after accounting for patient- and treatment-specific risk factors. In this proof of concept, 6 OAR DVH points and average monitor units serve as proxies for plan quality. Eighteen risk factors are considered for modeling quality: planning target volume (PTV) and OAR cross-sectional areas; volumes, spreads, and surface areas; minimum and centroid distances between OARs and the PTV; 6 PTV DVH points; use of chemotherapy; and surgery. A total of 69 head and neck cases were used to demonstrate the application of risk-adjusted control charts, and the results were compared with the application of conventional control charts. RESULTS: The risk-adjusted control chart remains robust to interpatient variations in the studied risk factors, unlike the conventional control chart. For the brainstem, the conventional chart signaled 4 patients with unusual (out-of-control) doses to 2% brainstem volume. However, the adjusted chart did not signal any plans after accounting for their risk factors. For the spinal cord doses to 2% brainstem volume, the conventional chart signaled 2 patients, and the adjusted chart signaled a separate patient after accounting for their risk factors. Similar adjustments were observed for the other DVH points when evaluating brainstem, spinal cord, ipsilateral parotid, and average monitor units. The adjustments can be directly attributed to the patient- and treatment-specific risk factors. CONCLUSIONS: A risk-adjusted control chart was developed to evaluate plan quality, which is robust to variations in patient- and treatment-specific parameters.