A deep learning-based prediction model for gamma evaluation in patient-specific quality assurance

Virtual Patient-Specific Quality Assurance of IMRT Using UNet++: Classification, Gamma Passing Rates Prediction, and Dose Difference Prediction

The dose verification in radiotherapy quality assurance (QA) is time-consuming and places a heavy workload on medical physicists. To provide a clinical tool to perform patient specific QA accurately, the UNet++ is investigated to classify failed or pass fields (the GPR lower than 85% is considered “failed” while the GPR higher than 85% is considered “pass”), predict gamma passing rates (GPR) for different gamma criteria, and predict dose difference from virtual patient-specific quality assurance in radiotherapy. UNet++ was trained and validated with 473 fields and tested with 95 fields. All plans used Portal Dosimetry for dose verification pre-treatment. Planar dose distribution of each field was used as the input for UNet++, with QA classification results, gamma passing rates of different gamma criteria, and dose difference were used as the output. In the test set, the accuracy of the classification model was 95.79%. The mean absolute error (MAE) were 0.82, 0.88, 2.11, 2.52, and the root mean squared error (RMSE) were 1.38, 1.57, 3.33, 3.72 for 3%/3mm, 3%/2 mm, 2%/3 mm, 2%/2 mm, respectively. The trend and position of the predicted dose difference were consistent with the measured dose difference. In conclusion, the Virtual QA based on UNet++ can be used to classify the field passed or not, predict gamma pass rate for different gamma criteria, and predict dose difference. The results show that UNet++ based Virtual QA is promising in quality assurance for radiotherapy.

Artificial intelligence for quality assurance in radiotherapy

In radiotherapy, patient-specific quality assurance is very time-consuming and causes machine downtime. It consists of testing (using measurement with a phantom and detector) if a modulated plan is correctly delivered by a treatment unit. Artificial intelligence and in particular machine learning algorithms were mentioned in recent reports as promising solutions to reduce or eliminate the patient-specific quality assurance workload. Several teams successfully experienced a virtual patient-specific quality assurance by training a machine learning tool to predict the results. Training data are generally composed of previous treatment plans and associated patient-specific quality assurance results. However, other training data types were recently introduced such as actual positions and velocities of multileaf collimators, metrics of the plan's complexity, and gravity vectors. Different types of machine learning algorithms were investigated (Poisson regression algorithms, convolutional neural networks, support vector classifiers) with sometimes promising results. These tools are being used for treatment units' quality assurance as well, in particular to analyse the results of imaging devices. Most of these reports were feasibility studies. Using machine learning in clinical routines as a tool that could fully replace quality assurance tests conducted by physics teams has yet to be implemented.

Commissioning and clinical implementation of an Autoencoder based Classification-Regression model for VMAT patient-specific QA in a multi-institution scenario

BACKGROUND AND PURPOSE

To commission and implement an Autoencoder based Classification-Regression (ACLR) model for VMAT patient-specific quality assurance (PSQA) in a multi-institution scenario.

MATERIALS AND METHODS

1835 VMAT plans from seven institutions were collected for the ACLR model commissioning and multi-institutional validation. We established three scenarios to validate the gamma passing rates (GPRs) prediction and classification accuracy with the ACLR model for different delivery equipment, QA devices, and treatment planning systems (TPS). The prediction performance of the ACLR model was evaluated using mean absolute error (MAE) and root mean square error (RMSE). The classification performance was evaluated using sensitivity and specificity. An independent end-to-end test (E2E) and routine QA of the ACLR model were performed to validate the clinical use of the model.

RESULTS

For multi-institution validations, the MAEs were 1.30-2.80% and 2.42-4.60% at 3%/3 mm and 3%/2 mm, respectively, and RMSEs were 1.55-2.98% and 2.83-4.95% at 3%/3 mm and 3%/2 mm, respectively, with different delivery equipment, QA devices, and TPS, while the sensitivity was 90% and specificity was 70.1% at 3%/2 mm. For the E2E, the deviations between the predicted and measured results were within 3%, and the model passed the consistency check for clinical implementation. The predicted results of the model were the same in daily QA, while the deviations between the repeated monthly measured GPRs were all within 2%.

CONCLUSIONS

The performance of the ACLR model in multi-institution scenarios was validated on a large scale. Routine QA of the ACLR model was established and the model could be used for VMAT PSQA clinically.

Error detection model developed using a multi-task convolutional neural network in patient-specific quality assurance for volumetric-modulated arc therapy

PURPOSE

In patient-specific quality assurance (QA) for static beam intensity-modulated radiation therapy (IMRT), machine-learning-based dose analysis methods have been developed to identify the cause of an error as an alternative to gamma analysis. Although these new methods have revealed that the cause of the error can be identified by analyzing the dose distribution obtained from the two-dimensional detector, they have not been extended to the analysis of volumetric-modulated arc therapy (VMAT) QA. In this study, we propose a deep learning approach to detect various types of errors in patient-specific VMAT QA.

METHODS

A total of 161 beams from 104 prostate VMAT plans were analyzed. All beams were measured using a cylindrical detector (Delta4; ScandiDos, Uppsala, Sweden), and predicted dose distributions in a cylindrical phantom were calculated using a treatment planning system (TPS). In addition to the error-free plan, we simulated 12 types of errors: two types of multileaf collimator positional errors (systematic or random leaf offset of 2 mm), two types of monitor unit (MU) scaling errors (±3%), two types of gantry rotation errors (±2° in clockwise and counterclockwise direction), and six types of phantom setup errors (±1 mm in lateral, longitudinal, and vertical directions). The error-introduced predicted dose distributions were created by editing the calculated dose distributions using a TPS with in-house software. Those 13 types of dose difference maps, consisting of an error-free map and 12 error maps, were created from the measured and predicted dose distributions and were used to train the convolutional neural network (CNN) model. Our model was a multi-task model that individually detected each of the 12 types of errors. Two datasets, Test sets 1 and 2, were prepared to evaluate the performance of the model. Test set 1 consisted of 13 types of dose maps used for training, whereas Test set 2 included the dose maps with 25 types of errors in addition to the error-free dose map. The dose map, which introduced 25 types of errors, was generated by combining two of the 12 types of simulated errors. For comparison with the performance of our model, gamma analysis was performed for Test sets 1 and 2 with the criteria set to 3%/2 mm and 2%/1 mm (dose difference/distance to agreement).

RESULTS

For Test set 1, the overall accuracy of our CNN model, gamma analysis with the criteria set to 3%/2 mm, and gamma analysis with the criteria set to 2%/1 mm was 0.92, 0.19, and 0.81, respectively. Similarly, for Test set 2, the overall accuracy was 0.44, 0.42, and 0.95, respectively. Our model outperformed gamma analysis in the classification of dose maps containing a single type error, and the performance of our model was inferior in the classification of dose maps containing compound errors.

CONCLUSIONS

A multi-task CNN model for detecting errors in patient-specific VMAT QA using a cylindrical measuring device was constructed, and its performance was evaluated. Our results demonstrate that our model was effective in identifying the error type in the dose map for VMAT QA.

Artificial Intelligence in magnetic Resonance guided Radiotherapy: Medical and physical considerations on state of art and future perspectives

Over the last years, technological innovation in Radiotherapy (RT) led to the introduction of Magnetic Resonance-guided RT (MRgRT) systems. Due to the higher soft tissue contrast compared to on-board CT-based systems, MRgRT is expected to significantly improve the treatment in many situations. MRgRT systems may extend the management of inter- and intra-fraction anatomical changes, offering the possibility of online adaptation of the dose distribution according to daily patient anatomy and to directly monitor tumor motion during treatment delivery by means of a continuous cine MR acquisition. Online adaptive treatments require a multidisciplinary and well-trained team, able to perform a series of operations in a safe, precise and fast manner while the patient is waiting on the treatment couch. Artificial Intelligence (AI) is expected to rapidly contribute to MRgRT, primarily by safely and efficiently automatising the various manual operations characterizing online adaptive treatments. Furthermore, AI is finding relevant applications in MRgRT in the fields of image segmentation, synthetic CT reconstruction, automatic (on-line) planning and the development of predictive models based on daily MRI. This review provides a comprehensive overview of the current AI integration in MRgRT from a medical physicist's perspective. Medical physicists are expected to be major actors in solving new tasks and in taking new responsibilities: their traditional role of guardians of the new technology implementation will change with increasing emphasis on the managing of AI tools, processes and advanced systems for imaging and data analysis, gradually replacing many repetitive manual tasks.

Current landscape and potential future applications of artificial intelligence in medical physics and radiotherapy

Artificial intelligence (AI) has seen tremendous growth over the past decade and stands to disrupts the medical industry. In medicine, this has been applied in medical imaging and other digitised medical disciplines, but in more traditional fields like medical physics, the adoption of AI is still at an early stage. Though AI is anticipated to be better than human in certain tasks, with the rapid growth of AI, there is increasing concerns for its usage. The focus of this paper is on the current landscape and potential future applications of artificial intelligence in medical physics and radiotherapy. Topics on AI for image acquisition, image segmentation, treatment delivery, quality assurance and outcome prediction will be explored as well as the interaction between human and AI. This will give insights into how we should approach and use the technology for enhancing the quality of clinical practice.

Systematic method for a deep learning-based prediction model for gamma evaluation in patient-specific quality assurance of volumetric modulated arc therapy

PURPOSE

This study aimed to develop and evaluate a novel strategy for establishing a deep learning-based gamma passing rate (GPR) prediction model for volumetric modulated arc therapy (VMAT) using dummy target plan data, one measurement process, and a multicriteria prediction method.

METHODS

A total of 147 VMAT plans were used for the training set (two sets of 48 dummy target plans) and test set (51 clinical target plans). The dummy plans were measured using a diode array detector. We developed an original convolutional neural network that accepts coronal and sagittal dose distributions to predict the GPRs of 36 pairs of gamma criteria from 0.5%/0.5 mm to 3%/3 mm. Sixfold cross-validation and model averaging were performed, and the mean training result and mean test result were derived from six trained models that were produced during cross-validation.

RESULTS

Strong or moderate correlations were observed between the measured and predicted GPRs in all criteria. The mean absolute errors and root mean squared errors of the test set (clinical target plan) were 0.63 and 1.11 in 3%/3 mm, 1.16 and 1.73 in 3%/2 mm, 1.96 and 2.66 in 2%/2 mm, 5.00 and 6.35 in 1%/1 mm, and 5.42 and 6.78 in 0.5%/1 mm, respectively. The Pearson correlation coefficients were 0.80 in the training set and 0.68 in the test set at the 0.5%/1 mm criterion.

CONCLUSION

Our results suggest that the training of the deep learning-based quality assurance model can be performed using a dummy target plan.

A tool for patient-specific prediction of delivery discrepancies in machine parameters using trajectory log files

PURPOSE

Multileaf collimator (MLC) delivery discrepancy between planned and actual (delivered) positions have detrimental effect on the accuracy of dose distributions for both IMRT and VMAT. In this study, we evaluated the consistency of MLC delivery discrepancies over the course of treatment and over time to verify that a predictive machine learning model would be applicable throughout the course of treatment. Next, the MLC and gantry positions recorded in prior trajectory log files were analyzed to build a machine learning algorithm to predict MLC positional discrepancies during delivery for a new treatment plan. An open source tool was developed and released to predict the MLC positional discrepancies at treatment delivery for any given plan.

METHODS

Trajectory log files of 142 IMRT plans and 125 VMAT plans from 9 Varian TrueBeam linear accelerators were collected and analyzed. The consistency of delivery discrepancy over patient-specific quality assurance (QA) and patient treatment deliveries was evaluated. Data were binned by treatment site and machine type to determine their relationship with MLC and gantry angle discrepancies. Motion-related parameters including MLC velocity, MLC acceleration, control point, dose rate, and gravity vector, gantry velocity and gantry acceleration, where applicable, were analyzed to evaluate correlations with MLC and gantry discrepancies. Several regression models, such as simple/multiple linear regression, decision tree, and ensemble method (boosted tree and bagged tree model) were used to develop a machine learning algorithm to predict MLC discrepancy based on MLC motion parameters.

RESULTS

MLC discrepancies at patient-specific QA differed from those at patient treatment deliveries by a small (mean = 0.0021 ± 0.0036 mm, P = 0.0089 for IMRT; mean = 0.0010 ± 0.0016 mm, P = 0.0003 for VMAT) but statistically significant amount, likely due to setting the gantry angle to zero for QA in IMRT. MLC motion parameters, MLC velocity and gravity vector, showed significant correlation (P < 0.001) with MLC discrepancy, especially MLC velocity, which had an approximately linear relationship (slope = -0.0027, P < 0.001, R²  = 0.79). Incorporating MLC motion parameters, the final generalized model trained by data from all linear accelerators can predict MLC discrepancy to a high degree of accuracy with high correlation (R²  = 0.86) between predicted and actual MLC discrepancies. The same prediction results were found across different treatment sites and linear accelerators.

CONCLUSION

We have developed a machine learning model using trajectory log files to predict the MLC discrepancies during delivery. This model has been a released as a research tool in which a DICOM-RT with predicted MLC positions can be generated using the original DICOM-RT file as input. This tool can be used to simulate radiotherapy treatment delivery and may be useful for studies evaluating plan robustness and dosimetric uncertainties from treatment delivery.

Auto-Trending daily quality assurance program for a pencil beam scanning proton system aligned with TG 224

The Daily Quality Assurance (DQA) for a proton modality is not standardized. The modern pencil beam scanning proton system is becoming a trend and an increasing number of proton centers with PBS are either under construction or in planning. The American Association of Physicists in Medicine has a Task Group 224 report published in 2019 for proton modality routine QA. Therefore, there is a clinical need to explore a DQA procedure to meet the TG 224 guideline. The MatriXX PT and a customized phantom were used for the dosimetry constancy checking. An OBI box was used for imaging QA. The MyQA(TM) software was used for logging the dosimetry results. An in‐house developed application was applied to log and auto analyze the DQA results. Another in‐house developed program "DailyQATrend" was used to create DQA databases for further analysis. All the functional and easy determined tasks passed. For dosimetry constancy checking, the outputs for four gantry rooms were within ±3% with room to room baseline differences within ±1%. The energy checking was within ±1%. The spot location checking from the baseline was within 0.63 mm and the spot size checking from the baseline was within −1.41 ± 1.27 mm (left–right) and −0.24 ± 1.27 mm (in–out) by averaging all the energies. We have found that there was also a trend for the beam energies of two treatment rooms slowly going down (0.76% per month and 0.48 per month) after analyzing the whole data trend with linear regression. A DQA program for a PBS proton system has been developed and fully implemented into the clinic. The DQA program meets the TG 224 guideline and has web‐based logging and auto treading functions. The clinical data show the DQA program is efficient and has the potential to identify the PBS proton system potential issue.

Dose Super-Resolution in Prostate Volumetric Modulated Arc Therapy Using Cascaded Deep Learning Networks

Purpose

This study proposes a cascaded network model for generating high-resolution doses (i.e., a 1 mm grid) from low-resolution doses (i.e., ≥3 mm grids) with reduced computation time.

Methods

Using the anisotropic analytical algorithm with three grid sizes (1, 3, and 5 mm) and the Acuros XB algorithm with two grid sizes (1 and 3 mm), dose distributions were calculated for volumetric modulated arc therapy plans for 73 prostate cancer patients. Our cascaded network model consisted of a hierarchically densely connected U-net (HD U-net) and a residual dense network (RDN), which were trained separately following a two-dimensional slice-by-slice procedure. The first network (HD U-net) predicted the downsampled high-resolution dose (generated through bicubic downsampling of the baseline high-resolution dose) using the low-resolution dose; subsequently, the second network (RDN) predicted the high-resolution dose from the output of the first network. Further, the predicted high-resolution dose was converted to its absolute value. We quantified the network performance using the spatial/dosimetric parameters (dice similarity coefficient, mean dose, maximum dose, minimum dose, homogeneity index, conformity index, and V_95%, V_70%, V_50%, and V_30%) for the low-resolution and predicted high-resolution doses relative to the baseline high-resolution dose. Gamma analysis (between the baseline dose and the low-resolution dose/predicted high-resolution dose) was performed with a 2%/2 mm criterion and 10% threshold.

Results

The average computation time to predict a high-resolution axial dose plane was <0.02 s. The dice similarity coefficient values for the predicted doses were closer to 1 when compared to those for the low-resolution doses. Most of the dosimetric parameters for the predicted doses agreed more closely with those for the baseline than for the low-resolution doses. In most of the parameters, no significant differences (p-value of >0.05) between the baseline and predicted doses were observed. The gamma passing rates for the predicted high-resolution does were higher than those for the low-resolution doses.

Conclusion

The proposed model accurately predicted high-resolution doses for the same dose calculation algorithm. Our model uses only dose data as the input without additional data, which provides advantages of convenience to user over other dose super-resolution methods.

Statistical Analysis of Treatment Planning Parameters for Prediction of Delivery Quality Assurance Failure for Helical Tomotherapy

Purpose:

This study aimed to investigate the parameters with a significant impact on delivery quality assurance (DQA) failure and analyze the planning parameters as possible predictors of DQA failure for helical tomotherapy.

Methods:

In total, 212 patients who passed or failed DQA measurements were retrospectively included in this study. Brain (n = 43), head and neck (n = 37), spinal (n = 12), prostate (n = 36), rectal (n = 36), pelvis (n = 13), cranial spinal irradiation and a treatment field including lymph nodes (n = 24), and other types of cancer (n = 11) were selected. The correlation between DQA results and treatment planning parameters were analyzed using logistic regression analysis. Receiver operating characteristic (ROC) curves, areas under the curves (AUCs), and the Classification and Regression Tree (CART) algorithm were used to analyze treatment planning parameters as possible predictors for DQA failure.

Results:

The AUC for leaf open time (LOT) was 0.70, and its cut-off point was approximately 30%. The ROC curve for the predicted probability calculated when the multivariate variable model was applied showed an AUC of 0.815. We confirmed that total monitor units, total dose, and LOT were significant predictors for DQA failure using the CART.

Conclusions:

The probability of DQA failure was higher when the percentage of LOT below 100 ms was higher than 30%. The percentage of LOT below 100 ms should be considered in the treatment planning process. The findings from this study may assist in the prediction of DQA failure in the future.

Multi-task autoencoder based classification-regression model for patient-specific VMAT QA

Patient-specific quality assurance (PSQA) of volumetric modulated arc therapy (VMAT) to assure accurate treatment delivery is resource-intensive and time-consuming. Recently, machine learning has been increasingly investigated in PSQA results prediction. However, the classification performance of models at different criteria needs further improvement and clinical validation (CV), especially for predicting plans with low gamma passing rates (GPRs). In this study, we developed and validated a novel multi-task model called autoencoder based classification-regression (ACLR) for VMAT PSQA. The classification and regression were integrated into one model, both parts were trained alternatively while minimizing a defined loss function. The classification was used as an intermediate result to improve the regression accuracy. Different tasks of GPRs prediction and classification based on different criteria were trained simultaneously. Balanced sampling techniques were used to improve the prediction accuracy and classification sensitivity for the unbalanced VMAT plans. Fifty-four metrics were selected as inputs to describe the plan modulation-complexity and delivery-characteristics, while the outputs were PSQA GPRs. A total of 426 clinically delivered VMAT plans were used for technical validation (TV), and another 150 VMAT plans were used for CV to evaluate the generalization performance of the model. The ACLR performance was compared with the Poisson Lasso (PL) model and found significant improvement in prediction accuracy. In TV, the absolute prediction error (APE) of ACLR was 1.76%, 2.60%, and 4.66% at 3%/3 mm, 3%/2 mm, and 2%/2 mm, respectively; whereas the APE of PL was 2.10%, 3.04%, and 5.29% at 3%/3 mm, 3%/2 mm, and 2%/2 mm, respectively. No significant difference was found between CV and TV in prediction accuracy. ACLR model set with 3%/3 mm can achieve 100% sensitivity and 83% specificity. The ACLR model could classify the unbalanced VMAT QA results accurately, and it can be readily applied in clinical practice for virtual VMAT QA.

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

PURPOSE

This study aims to investigate the use of machine learning models for delivery error prediction in proton pencil beam scanning (PBS) delivery.

METHODS

A dataset of planned and delivered PBS spot parameters was generated from a set of 20 prostate patient treatments. Planned spot parameters (spot position, MU and energy) were extracted from the treatment planning system (TPS) for each beam. Delivered spot parameters were extracted from irradiation log-files for each beam delivery following treatment. The dataset was used as a training dataset for three machine learning models which were trained to predict delivered spot parameters based on planned parameters. K-fold cross validation was employed for hyper-parameter tuning and model selection where the mean absolute error (MAE) was used as the model evaluation metric. The model with lowest MAE was then selected to generate a predicted dose distribution for a test prostate patient within a commercial TPS.

RESULTS

Analysis of the spot position delivery error between planned and delivered values resulted in standard deviations of 0.39 mm and 0.44 mm for x and y spot positions respectively. Prediction error standard deviation values of spot positions using the selected model were 0.22 mm and 0.11 mm for x and y spot positions respectively. Finally, a three-way comparison of dose distributions and DVH values for select OARs indicates that the random-forest-predicted dose distribution within the test prostate patient was in closer agreement to the delivered dose distribution than the planned distribution.

CONCLUSIONS

PBS delivery error can be accurately predicted using machine learning techniques.

Automated Intensity Modulated Radiation Therapy Treatment Planning for Cervical Cancer Based on Convolution Neural Network

Purpose:

To develop and evaluate an automatic intensity-modulated radiation therapy (IMRT) program for cervical cancer, including a Convolution Neural Network (CNN)-based prediction model and an automated optimization strategy.

Methods:

A CNN deep learning model was trained to predict a patient-specify set of IMRT objectives based on overlap volume histograms (OVH) and high-quality plan of previous patients. A total of 140 cervical cancer patients were enrolled in this study, including 100 patients in the training set, 20 patients in the validation set and 20 patients in the testing set. The input of this model was OVH data and the output were values of IMRT plan objectives. For patients in the testing set, the set of planning objectives were predicted by the CNN model and used to automatically generate IMRT plans. Meanwhile, manual plans of these patients were generated by 1 beginner planner and 1 senior planner respectively. Finally, dose distribution, dosimetric parameters and planning time were analyzed. In addition, the 3 types of plans were blinded compared and ranked by an experienced oncologist.

Results:

There were almost no statistically differences among these 3 types of plans in target coverage and dose conformity. Dose homogeneity were slightly decreased while the average dose and parameters for most organs-at-risk (OARs) were decreased in automatic plans. Especially in comparison with manual plans by the beginner planner, V₄₀ of bladder and rectum decreased 6.3% and 12.3%, while mean dose of rectum and marrow were 1.1 Gy and 1.8 Gy lower with automatic plans (either P < 0.017). In the blinded comparison, automatic plans were chosen as best plan in 14 cases.

Conclusions:

For cervical cancer, automatic IMRT plans optimized from the CNN generated objectives have superior dose sparing without compromising of target dose. It significantly reduced the planning time.

Identification of treatment error types for lung cancer patients using convolutional neural networks and EPID dosimetry

BACKGROUND/PURPOSE

Electronic portal imaging device (EPID) dosimetry aims to detect treatment errors, potentially leading to treatment adaptation. Clinically used threshold classification methods for detecting errors lead to loss of information (from multi-dimensional EPID data to a few numbers) and cannot be used for identifying causes of errors. Advanced classification methods, such as deep learning, can use all available information. In this study, convolutional neural networks (CNNs) were trained to detect and identify error type and magnitude of simulated treatment errors in lung cancer patients. The purpose of this simulation study is to provide a proof-of-concept of CNNs for error identification using EPID dosimetry in an in vivo scenario.

MATERIALS AND METHODS

Clinically realistic ranges of anatomical changes, positioning errors and mechanical errors were simulated for lung cancer patients. Predicted portal dose images (PDIs) containing errors were compared to error-free PDIs using the widely used gamma analysis. CNNs were trained to classify errors using 2D gamma maps. Three classification levels were assessed: Level 1 (main error type, e.g., anatomical change), Level 2 (error subtype, e.g., tumor regression) and Level 3 (error magnitude, e.g., >50% tumor regression).

RESULTS

CNNs showed good performance for all classification levels (training/test accuracy 99.5%/96.1%, 92.5%/86.8%, 82.0%/72.9%). For Level 3, overfitting became more apparent.

CONCLUSION

This simulation study indicates that deep learning is a promising powerful tool for identifying types and magnitude of treatment errors with EPID dosimetry, providing additional information not currently available from EPID dosimetry. This is a first step towards rapid, automated models for identification of treatment errors using EPID dosimetry.

Verification of the machine delivery parameters of a treatment plan via deep learning

We developed a generative adversarial network (GAN)-based deep learning approach to estimate the multileaf collimator (MLC) aperture and corresponding monitor units (MUs) from a given 3D dose distribution. The proposed design of the adversarial network, which integrates a residual block into pix2pix framework, jointly trains a 'U-Net'-like architecture as the generator and a convolutional 'PatchGAN' classifier as the discriminator. 199 patients, including nasopharyngeal, lung and rectum, treated with intensity-modulated radiotherapy and volumetric-modulated arc therapy techniques were utilized to train the network. An additional 47 patients were used to test the prediction accuracy of the proposed deep learning model. The Dice similarity coefficient (DSC) was calculated to evaluate the similarity between the MLC aperture shapes obtained from the treatment planning system (TPS) and the deep learning prediction. The average and standard deviation of the bias between the TPS-generated MUs and predicted MUs was calculated to evaluate the MU prediction accuracy. In addition, the differences between TPS and deep learning-predicted MLC leaf positions were compared. The average and standard deviation of DSC was 0.94 ± 0.043 for 47 testing patients. The average deviation of predicted MUs from the planned MUs normalized to each beam or arc was within 2% for all the testing patients. The average deviation of the predicted MLC leaf positions was around one pixel for all the testing patients. Our results demonstrated the feasibility and reliability of the proposed approach. The proposed technique has strong potential to improve the efficiency and accuracy of the patient plan quality assurance process.

Integration of AI and Machine Learning in Radiotherapy QA

The use of machine learning and other sophisticated models to aid in prediction and decision making has become widely popular across a breadth of disciplines. Within the greater diagnostic radiology, radiation oncology, and medical physics communities promising work is being performed in tissue classification and cancer staging, outcome prediction, automated segmentation, treatment planning, and quality assurance as well as other areas. In this article, machine learning approaches are explored, highlighting specific applications in machine and patient-specific quality assurance (QA). Machine learning can analyze multiple elements of a delivery system on its performance over time including the multileaf collimator (MLC), imaging system, mechanical and dosimetric parameters. Virtual Intensity-Modulated Radiation Therapy (IMRT) QA can predict passing rates using different measurement techniques, different treatment planning systems, and different treatment delivery machines across multiple institutions. Prediction of QA passing rates and other metrics can have profound implications on the current IMRT process. Here we cover general concepts of machine learning in dosimetry and various methods used in virtual IMRT QA, as well as their clinical applications.

Overview of artificial intelligence-based applications in radiotherapy: Recommendations for implementation and quality assurance

Artificial Intelligence (AI) is currently being introduced into different domains, including medicine. Specifically in radiation oncology, machine learning models allow automation and optimization of the workflow. A lack of knowledge and interpretation of these AI models can hold back wide-spread and full deployment into clinical practice. To facilitate the integration of AI models in the radiotherapy workflow, generally applicable recommendations on implementation and quality assurance (QA) of AI models are presented. For commonly used applications in radiotherapy such as auto-segmentation, automated treatment planning and synthetic computed tomography (sCT) the basic concepts are discussed in depth. Emphasis is put on the commissioning, implementation and case-specific and routine QA of AI models needed for a methodical introduction in clinical practice.

Artificial intelligence in radiation oncology

Artificial intelligence (AI) has the potential to fundamentally alter the way medicine is practised. AI platforms excel in recognizing complex patterns in medical data and provide a quantitative, rather than purely qualitative, assessment of clinical conditions. Accordingly, AI could have particularly transformative applications in radiation oncology given the multifaceted and highly technical nature of this field of medicine with a heavy reliance on digital data processing and computer software. Indeed, AI has the potential to improve the accuracy, precision, efficiency and overall quality of radiation therapy for patients with cancer. In this Perspective, we first provide a general description of AI methods, followed by a high-level overview of the radiation therapy workflow with discussion of the implications that AI is likely to have on each step of this process. Finally, we describe the challenges associated with the clinical development and implementation of AI platforms in radiation oncology and provide our perspective on how these platforms might change the roles of radiotherapy medical professionals.

Artificial intelligence in radiotherapy: a technological review

Radiation therapy (RT) is widely used to treat cancer. Technological advances in RT have occurred in the past 30 years. These advances, such as three-dimensional image guidance, intensity modulation, and robotics, created challenges and opportunities for the next breakthrough, in which artificial intelligence (AI) will possibly play important roles. AI will replace certain repetitive and labor-intensive tasks and improve the accuracy and consistency of others, particularly those with increased complexity because of technological advances. The improvement in efficiency and consistency is important to manage the increasing cancer patient burden to the society. Furthermore, AI may provide new functionalities that facilitate satisfactory RT. The functionalities include superior images for real-time intervention and adaptive and personalized RT. AI may effectively synthesize and analyze big data for such purposes. This review describes the RT workflow and identifies areas, including imaging, treatment planning, quality assurance, and outcome prediction, that benefit from AI. This review primarily focuses on deep-learning techniques, although conventional machine-learning techniques are also mentioned.

Improvement of prediction and classification performance for gamma passing rate by using plan complexity and dosiomics features

PURPOSE

The purpose of this study was to predict and classify the gamma passing rate (GPR) value by using new features (3D dosiomics features and combined with plan and dosiomics features) together with a machine learning technique for volumetric modulated arc therapy (VMAT) treatment plans.

METHODS AND MATERIALS

A total of 888 patients who underwent VMAT were enrolled comprising 1255 treatment plans. Further, 24 plan complexity features and 851 dosiomics features were extracted from the treatment plans. The dataset was randomly split into a training/validation (80%) and test (20%) dataset. The three models for prediction and classification using XGBoost were as follows: (i) plan complexity features-based prediction method (plan model); (ii) 3D dosiomics feature-based prediction model (dosiomics model); (iii) a combination of both the previous models (hybrid model). The prediction performance was evaluated by calculating the mean absolute error (MAE) and the correlation coefficient (CC) between the predicted and measured GPRs. The classification performance was evaluated by calculating the area under curve (AUC) and sensitivity.

RESULTS

MAE and CC at γ2%/2 mm in the test dataset were 4.6% and 0.58, 4.3% and 0.61, and 4.2% and 0.63 for the plan model, dosiomics model, and hybrid model, respectively. AUC and sensitivity at γ2%/2 mm in test dataset were 0.73 and 0.70, 0.81 and 0.90, and 0.83 and 0.90 for the plan model, dosiomics model, and hybrid model, respectively.

CONCLUSIONS

A combination of both plan and dosiomics features with machine learning technique can improve the prediction and classification performance for GPR.

An introduction to deep learning in medical physics: advantages, potential, and challenges

As one of the most popular approaches in artificial intelligence, deep learning (DL) has attracted a lot of attention in the medical physics field over the past few years. The goals of this topical review article are twofold. First, we will provide an overview of the method to medical physics researchers interested in DL to help them start the endeavor. Second, we will give in-depth discussions on the DL technology to make researchers aware of its potential challenges and possible solutions. As such, we divide the article into two major parts. The first part introduces general concepts and principles of DL and summarizes major research resources, such as computational tools and databases. The second part discusses challenges faced by DL, present available methods to mitigate some of these challenges, as well as our recommendations.

Impact of delivery characteristics on dose delivery accuracy of volumetric modulated arc therapy for different treatment sites

This study aimed to investigate the impact of delivery characteristics on the dose delivery accuracy of volumetric modulated arc therapy (VMAT) for different treatment sites. The pretreatment quality assurance (QA) results of 344 VMAT patients diagnosed with gynecological (GYN), head and neck (H&N), rectal or prostate cancer were randomly chosen in this study. Ten metrics reflecting VMAT delivery characteristics were extracted from the QA plans. Compared with GYN and rectal plans, H&N and prostate plans had higher aperture complexity and monitor units (MU), and smaller aperture area. Prostate plans had the smallest aperture area and lowest leaf speed compared with other plans (P < 0.001). No differences in gantry speed were found among the four sites. The gamma passing rates (GPRs) of GYN, rectal and H&N plans were inversely associated with union aperture area (UAA) and leaf speed (Pearson's r: -0.39 to -0.68). GPRs of prostate plans were inversely correlated with aperture complexity, MU and small aperture score (SAS) (absolute Pearson's r: 0.34 to 0.49). Significant differences in GPR between high SAS and low SAS subgroups were found only when leaf speed was <0.42 cm s-1 (P < 0.001). No association of GPR with gantry speed was found in four sites. Leaf speed was more strongly associated with UAA. Aperture complexity and MU were more strongly associated with SAS. VMAT plans from different sites have distinct delivery characteristics. Affecting dose delivery accuracy, leaf speed is the key factor for GYN, rectal and H&N plans, while aperture complexity, MU and small apertures have a higher influence on prostate plans.

Complexity metrics for IMRT and VMAT plans: a review of current literature and applications

Modulated radiotherapy with multileaf collimators is widely used to improve target conformity and normal tissue sparing. This introduced an additional degree of complexity, studied by multiple teams through different properties. Three categories of complexity metrics were considered in this review: fluence, deliverability and accuracy metrics. The first part of this review is dedicated to the inventory of these complexity metrics. Different applications of these metrics emerged. Influencing the optimizer by integrating complexity metrics into the cost function has been little explored and requires more investigations. In modern treatment planning system, it remains confined to MUs or treatment time limitation. A large majority of studies calculated metrics only for analysis, without plan modification. The main application was to streamline the patient specific quality assurance workload, investigating the capability of complexity metrics to predict patient specific quality assurance results. Additionally complexity metrics were used to analyze behaviour of TPS optimizer, compare TPS, operators and plan properties, and perform multicentre audit. Their potential was also explored in the context of adaptive radiotherapy and automation planning. The second part of the review gives an overview of these studies based on the complexity metrics.

Machine Learning for Patient-Specific Quality Assurance of VMAT: Prediction and Classification Accuracy

PURPOSE

To assess the accuracy of machine learning to predict and classify quality assurance (QA) results for volumetric modulated arc therapy (VMAT) plans.

METHODS AND MATERIALS

Three hundred three VMAT plans, including 176 gynecologic cancer and 127 head and neck cancer plans, were chosen in this study. Fifty-four complexity metrics were extracted from the QA plans and considered as inputs. Patient-specific QA was performed, and gamma passing rates (GPRs) were used as outputs. One Poisson lasso (PL) regression model was developed, aiming to predict individual GPR, and 1 random forest (RF) classification model was developed to classify QA results as "pass" or "fail." Both technical validation (TV) and clinical validation (CV) were used to evaluate the model reliability. GPR prediction accuracy of PL and classification performance of PL and RF were evaluated.

RESULTS

In TV, the mean prediction error of PL was 1.81%, 2.39%, and 4.18% at 3%/3 mm, 3%/2 mm, and 2%/2 mm, respectively. No significant differences in prediction errors between TV and CV were observed. In QA results classification, PL had a higher specificity (accurately identifying plans that can pass QA), whereas RF had a higher sensitivity (accurately identifying plans that may fail QA). By using 90% as the action limit at a 3%/2 mm criterion, the specificity of PL and RF was 97.5% and 87.7% in TV and 100% and 71.4% in CV, respectively. The sensitivity of PL and RF was 31.6% and 100% in TV and 33.3% and 100% in CV, respectively. With 100% sensitivity, the QA workload of 81.2% of plans in TV and 62.5% of plans in CV could be reduced by RF.

CONCLUSIONS

The PL model could accurately predict GPR for most VMAT plans. The RF model with 100% sensitivity was preferred for QA results classification. Machine learning can be a useful tool to assist VMAT QA and reduce QA workload.

Prediction of dosimetric accuracy for VMAT plans using plan complexity parameters via machine learning

PURPOSE

The dosimetric accuracies of volumetric modulated arc therapy (VMAT) plans were predicted using plan complexity parameters via machine learning.

METHODS

The dataset consisted of 600 cases of clinical VMAT plans from a single institution. The predictor variables (n = 28) for each plan included complexity parameters, machine type, and photon beam energy. Dosimetric measurements were performed using a helical diode array (ArcCHECK), and the dosimetric accuracy of the passing rates for a 5% dose difference (DD5%) and gamma index of 3%/3 mm (γ3%/3 mm) were predicted using three machine learning models: regression tree analysis (RTA), multiple regression analysis (MRA), and neural networks (NNs). First, the prediction models were applied to 500 cases of the VMAT plans. Then, the dosimetric accuracy was predicted using each model for the remaining 100 cases (evaluation dataset). The error between the predicted and measured passing rates was evaluated.

RESULTS

For the 600 cases, the mean ± standard deviation of the measured passing rates was 92.3% ± 9.1% and 96.8% ± 3.1% for DD5% and γ3%/3 mm, respectively. For the evaluation dataset, the mean ± standard deviation of the prediction errors for DD5% and γ3%/3 mm was 0.5% ± 3.0% and 0.6% ± 2.4% for RTA, 0.0% ± 2.9% and 0.5% ± 2.4% for MRA, and -0.2% ± 2.7% and -0.2% ± 2.1% for NN, respectively.

CONCLUSIONS

NNs performed slightly better than RTA and MRA in terms of prediction error. These findings may contribute to increasing the efficiency of patient-specific quality-assurance procedures.

Diagnosis of Alzheimer's Disease via Multi-Modality 3D Convolutional Neural Network

Alzheimer's disease (AD) is one of the most common neurodegenerative diseases. In the last decade, studies on AD diagnosis has attached great significance to artificial intelligence-based diagnostic algorithms. Among the diverse modalities of imaging data, T1-weighted MR and FDG-PET are widely used for this task. In this paper, we propose a convolutional neural network (CNN) to integrate all the multi-modality information included in both T1-MR and FDG-PET images of the hippocampal area, for the diagnosis of AD. Different from the traditional machine learning algorithms, this method does not require manually extracted features, instead, it utilizes 3D image-processing CNNs to learn features for the diagnosis or prognosis of AD. To test the performance of the proposed network, we trained the classifier with paired T1-MR and FDG-PET images in the ADNI datasets, including 731 cognitively unimpaired (labeled as CN) subjects, 647 subjects with AD, 441 subjects with stable mild cognitive impairment (sMCI) and 326 subjects with progressive mild cognitive impairment (pMCI). We obtained higher accuracies of 90.10% for CN vs. AD task, 87.46% for CN vs. pMCI task, and 76.90% for sMCI vs. pMCI task. The proposed framework yields a state-of-the-art performance. Finally, the results have demonstrated that (1) segmentation is not a prerequisite when using a CNN for the classification, (2) the combination of two modality imaging data generates better results.

Predicting VMAT patient-specific QA results using a support vector classifier trained on treatment plan characteristics and linac QC metrics

The use of treatment plan characteristics to predict patient-specific quality assurance (QA) measurement results has recently been reported as a strategy to help facilitate automated pre-treatment verification workflows or to provide a virtual assessment of delivery quality. The goal of this work is to investigate the potential of using treatment plan characteristics and linac performance metrics (i.e. quality control test results) in combination with machine learning techniques to predict the results of VMAT patient-specific QA measurements. Using features that describe treatment plan complexity and linac performance metrics, we trained a linear support vector classifier (SVC) to classify the results of VMAT patient-specific QA measurements. The 'targets' in this model were simple classes representing median dose difference between measured and expected dose distributions-'hot' if the median dose deviation was  >1%, 'cold' if it was  <-1%, and 'normal' if it was within  ±1%. A total of 1620 unique patient-specific QA measurements were available for model development and testing. 75% of the data were used to develop and cross-validate the model, and the remaining 25% were used for an independent assessment of model performance. For the model development phase, a recursive feature elimination (RFE) cross-validation technique was used to eliminate unimportant features. Model performance was assessed using receiver operator characteristic (ROC) curve metrics. Of the ten features found to be most predictive of patient-specific QA measurement results, half were derived from treatment plan characteristics and half from quality control (QC) metrics characterizing linac performance. The model achieved a micro-averaged area under the ROC curve of 0.93, and a macro-averaged area under the ROC curve of 0.88. This work demonstrates the potential of using both treatment plan characteristics and routine linac QC results in the development of machine learning models for VMAT patient-specific QA measurements.