The success of modern radiotherapy treatment depends on the correct alignment of the radiation beams with the target region in the patient. In the conventional paradigm of image-guided radiation therapy, 2D or 3D setup images are taken immediately prior to treatment and are used by radiation therapy technologists to localize the patient to the same position as defined from the reference planning CT dataset. However, numerous reports in the literature have described errors during this step, which have led to incorrect treatments and potentially significant clinical harm to patients. In addition, reported errors likely underestimate the true error rate, as many errors may pass by undetected or are simply not reported. The human factor has been shown to play a large role in these errors, where the setup and planning CT imaging registration is not interpreted or performed correctly as per standard practice.

The hypothesis of the proposed study was to develop a workflow that can algorithmically compare 3D setup and planning CT imaging using image similarity metrics. The proposed system, intended to work in an automated and real-time fashion immediately prior to radiotherapy delivery, has the potential to act as a robust second-check safety interlock to prevent any identification or misalignment errors from reaching the patient. As no additional equipment is required in the treatment room or for patient setup, this system adds virtually no additional complexity, time, or cost to the treatment process. It can be applicable to countries around the world and is particularly relevant for developing nations, where higher error rates have been reported.

We examined both 3D cone beam CT and 3D megavoltage CT images covering the head-and-neck, spine, and pelvis regions. Workflows were developed for detecting errors in patient identification and patient misalignment, including various pre-processing and feature selection steps. Classification models were trained and evaluated by comparing misclassification errors, ROC curves, likelihood ratios, and more. Our system can achieve high accuracy and sensitivity/specificity rates for error detection in both patient identification and alignment scenarios, and there are many possible avenues to further test and improve the system’s robustness.

The research aims presented in this dissertation are centered on broad themes of improving automation and error detection in radiation oncology. The first aim was to create a stereoscopic radiographic generator for the ExacTrac image-guidance system. Our methodology enables medical physicists to compute geometric parameters for ray tracing based on values contained in ExacTrac configuration logs. Rigid registrations to ground-truth radiographs were performed using medical imaging software, and the results demonstrated sub-millimeter accuracy.

The second aim was to create deep-learning models to automatically detect off-by-one vertebral body misalignments with on-board planar imaging. Thoracic and abdominal radiotherapy plans were retrieved from our clinical servers using the DICOM networking protocol. Pairs of digital and treatment radiographs were organized according to beam energy and orientation. Realistic off-by-one misalignments were systematically produced. Convolutional neural networks were trained to classify whether such radiographic pairs are aligned or misaligned. Given a desired 95% model specificity, the orthogonal kilovoltage model achieved a sensitivity of 99%. The models established an independent review process for setup error incidents over large retrospective datasets, which are nearly impossible to review by hand. Of particular emphasis, an instance of a previously unnoticed off-by-one setup error at our institution was found.

The third aim was to create automated algorithms to evaluate the quality of prostate radiotherapy treatment plans. Quality metrics included number of days to plan approval, target margins, presence of fiducial markers, and prescribed radiation dose. The automated measurements were compared with values determined manually in clinical software, and high accuracy was obtained. Furthermore, deep-learning models for auto-contouring prostate bed target volumes were created. Refined models were developed using a novel data-driven approach to separate contours based on anterior and posterior convexity. Quality outliers were flagged for retrospective human review. Among the cases reviewed, a previously unnoticed mistake was identified, where a prostate patient treated at our institution was overexposed by 2 Gy. The automated algorithms were applied on treatment plans from our institution and from hospitals in the greater community, which allowed us to assess the existing range of standards of care in clinical practice.

Anatomical motion during external beam radiotherapy can reduce the accuracy of radiation delivery and degrade treatment efficacy. Magnetic resonance image (MRI) guided radiotherapy systems provide unparalleled soft tissue contrast and the opportunity to monitor tumor motion in images acquired during treatment. Radiotherapy gating, turning the radiation beam on when the target is within the desired spatial window and off when the target moves outside this window, is performed to mitigate the effects of anatomical motion. A limitation of image-based respiratory gating is the requirement that the images must be acquired rapidly and continuously to avoid unacceptable gating latency. Consequently, images are typically acquired at a single slice position using a rapid imaging sequence. Pulse sequence selection and thereby image contrast is limited to those available using rapid imaging sequences. Only the portion of the target within the imaged slice can be tracked and target motion outside the currently imaged slice is not evaluated. Additionally, due to electronic, mechanical and computational constraints, MRI-guided radiotherapy systems cannot respond instantaneously to anatomical motion. System latencies can reduce the radiation dose delivered to the target and increase the radiation delivered to nearby normal tissues.

As a preliminary step, we evaluate characteristics of the MRI-guided radiotherapy machine at UCLA relevant to intra-fraction motion management. Specifically, we evaluate the geometric distortion and dosimetric accuracy of the ViewRay MRIdian MRI-guided radiotherapy system. We measure geometric distortion using a commercially available phantom and software developed in our laboratory. We measure the dosimetric accuracy of gated radiotherapy delivery using a motion phantom and radiochromic film.

The key contribution of our work is to develop motion model based methods to overcome limitations of the current technology arising from finite imaging speeds and beam triggering latencies. We propose the use of a motion model and external respiratory bellows surrogate to allow pulse sequences that require significant delays between image acquisitions to be used for radiotherapy gating. Additionally, we develop motion modeling methods to estimate tumor motion across multiple slice positions in real-time to inform radiotherapy gating decisions and provide three-dimensional target visualization. To overcome system latencies and improve the current clinical deformable registration-based target tracking algorithm we develop a novel motion prediction method. Our method exploits the rich image intensity information available during real-time MRI and can predict motion across the entire field of view in real-time. Finally, we propose an accompanying framework to evaluate motion prediction confidence to detect when predictions are likely to be inaccurate.

Although safety procedures are in place within the radiation therapy (RT) workflow, incidents are still occurring due to human errors. To enhance patient safety, it is critical to identify and consolidate the vulnerabilities present within the various processes performed prior and during treatment. One such domain is within the Cone Beam Computed Tomography (CBCT)-guided RT workflow where there is currently no built-in system to check for errors in the registration of the simulation Computed Tomography (simCT) to the setup CBCT performed during the patient positioning step prior to radiation beam delivery. This lack of safeguards poses a risk to the patient as an incorrect registration may go undetected, leading to a compromised patient setup and treatment. The overall objective of the proposed work is to develop a deep learning-based error detection algorithm (EDA) which can serve as a secondary safety check to the radiation therapist while also helping to consolidate the robustness of CBCT-guided radiotherapy treatments. Additionally, this work explores the feasibility of a fully-unsupervised anomaly detection framework (ADF), based on a CBCT inpainting technique using a variational autoencoder, which would highlight anomalies for human review during regular physics quality assurance chart checks.

Initial results show that the EDA has a strong error-catching ability with areas under the receiver operating characteristic (ROC) curves of at least 99.2% when tested on simulated translational errors. When assessed against expert observers in a qualitative assessment of patient setup registrations, the EDA’s predictions achieved statistically significant correlations to the observer scores. Additionally, during a retrospective error search on a multi-institutional dataset of 17,612 registrations, the EDA successfully flagged the three known patient-setup incidents, and additionally identified four previously unreported incidents, proving its effectiveness on real-life cases. Those results validated the clinical utility of the EDA for bulk image reviews and highlighted the reliability and safety of CBCT-guided RT, with an absolute gross patient misalignment error rate of 0.04% ± 0.02% per delivered fraction. The ADF also demonstrated promising error detection ability when applied to a test dataset containing real patient setup incidents and simulated translational errors, with an area under the ROC curve of 98.1%.

The results described in this work validate the clinical utility and strong error-catching ability of both the EDA and the ADF when applied to real-world cases. We demonstrated that EDA and ADF can facilitate bulk image reviews, which can be useful for incident learning and can also expedite regular quality assurance chart checks performed by physicists or physicians. Additionally, if applied in real-time, EDA can consolidate the safety of CBCT-guided radiotherapy by serving as a secondary safety check to the therapist, thereby minimizing the risk of gross patient setup errors. Whether used prospectively or retrospectively, we believe that the proposed tools can add substantial value to the safety aspect of radiotherapy treatments, especially in low-middle income communities where the lack of workforce and safeguards often translates into a higher risk of treatment incidents.

The introduction of image guided radiation therapy (IGRT), defined as regular patient imaging during the course of a radiation therapy treatment regimen, has resulted in significant improvements in both the quality and safety of radiotherapy treatments. Accurate visualization of the target and nearby normal tissue has allowed for the reduction of planning target volume (PTV) margins, leading to increased normal tissue sparing. Daily image guidance can also increase the safety of radiation therapy treatments, as the patient is imaged prior to each treatment fraction to verify target location and compensate for inter-fraction anatomical changes. However, the introduction of daily imaging into the clinical workflow has been coupled with a simultaneous introduction of so-called “IGRT errors.” IGRT errors arise from inaccuracies in registering the patient’s daily setup images with the simulation images acquired prior to treatment. Such errors could arise due to technical challenges with the image registration algorithms themselves, problems with applying the image registration algorithms to a particular set of patient images, or human mistakes made while interpreting the results of an image registration. While IGRT has increased the precision of radiotherapy treatments, it can lead to treatments that are “precisely wrong.” As a preliminary step to mitigate IGRT errors, we propose the development of novel tools for the automatic detection of IGRT errors. Specifically, we develop a convolutional neural network (CNN)-based model for detecting the rare but serious IGRT error of off-by-one vertebral body misalignments in radiation therapy treatments targeting the thoracic spine. We develop a second CNN-based model for detecting the more generic IGRT error of translational shifts of 1 cm from treatment isocenter in all anatomic regions. We apply both models to retrospective image data from patients aligned using daily image guidance in order to detect previously unreported IGRT errors and near miss events, and to understand where in the clinical workflow such incidents originated. Finally, we understand that new evidence-based tools can only be effective if they are successfully integrated into the clinical environment. A rigorous implementation science approach is a necessary step to integrating novel technologies and reducing the well-documented lag time from research to practice. We study the barriers and facilitators to use of both automated tools that are commercially available as well as automated tools still in development. We use a survey study to evaluate medical dosimetrists’ perceptions of auto-contouring and automated treatment planning tools and their perceived barriers to regular clinical use of such tools. To better understand how a new automated tool designed to assist in the IGRT review portion of weekly chart checks could be integrated clinically, we use a novel thematic analysis approach to analyze the current weekly chart check workflow from the perspective of the clinical medical physicist.

Background and purpose

Materials and methods

Results

Conclusion

Purpose

Methods

Results

Conclusion

Purpose

Methods

Results

Conclusion

Purpose

Methods

Results

Conclusions

Purpose

Methods and materials

Results

Conclusions