Author ORCID Identifier
Date of Graduation
5-2025
Document Type
Dissertation (PhD)
Program Affiliation
Medical Physics
Degree Name
Doctor of Philosophy (PhD)
Advisor/Committee Chair
Laurence E. Court
Committee Member
David T. Fuentes
Committee Member
Anuja Jhingran
Committee Member
Christine B. Peterson
Committee Member
Julianne Pollard-Larkin
Committee Member
Surendra Prajapati
Committee Member
Thomas J. Whitaker
Abstract
Peer review of organ-at-risk and target volume delineation is essential for patient safety and optimization of treatment outcomes. Peer review makes a substantial impact on patient outcomes, reporting notable rates of plan changes when reviewed and worse survival when patient cases are not reviewed. In around half of revised cases, target volume change is the required cause for modification (related to tumor control); in one in ten cases, it is normal tissue sparing (related to treatment toxicity). Essentially all North American institutions with accredited residency training programs hold peer review to some capacity. However, accessibility to practicing routine peer review (especially of contours) is highly dependent on the presence of sub-specialized radiation oncology staff and their expensive time, making the workflow and qualitative decision-making inconsistent across all clinical practices, particularly those in low-resource settings. In pursuing the automation of the contour review process, especially with the wide adoption of auto-contouring to clinical practice, we aim to mitigate time and specialty resource shortages to this critical clinical practice.
The purpose of this study was to develop automated tools that can accurately identify sub-optimal contours (inadequate target coverage, sub-optimal delineation of organs at risk to reduce toxicity), agnostic of their manual or automated generation. To accomplish this, we first examined the relationship between geometric and dosimetric agreement metrics amongst normal tissue auto-contours and clinical contours in discerning whether the current state of automated contour QA (which mainly relies on geometric comparisons) was appropriately defining contour errors and deploying effective metrics for error detection. We tested this relationship over a substantial dataset of head and neck patients. Using this knowledge, we further examined the necessity of incorporating dose-based comparisons into normal tissue contour QA to improve the detection of clinically significant errors. A two-contour QA system was developed, leveraging an independent auto-contouring system to cross-validate a primary auto-contouring system in its performance against clinical manual delineations. We employed a logistic regression model as a means by which geometric and dosimetric comparisons in a two-contour QA system can flag potential errors in a primary auto-contouring system compared to the clinical ground truth. Finally, we explored automated contour QA of target contours by employing statistical process control to explore an interpretable and scalable approach for continuous contour quality monitoring of internal target volumes in cervical cancer. Control charts were generated to establish data-driven control limits for geometric and dosimetric comparisons between auto- and clinical target contours, allowing for automated flagging of deviations from expected contouring norms. This dissertation advances automated contour QA by integrating geometric, dosimetric, machine learning, and statistical process control techniques with auto-contouring, providing a framework for artificial intelligence-assisted contour verification that is interpretable, scalable, and clinically relevant. The tools developed in this work are expected to improve real-time contour QA through implementation to a web-based automated, expert peer review system, enabling clinics to practice contour peer review despite disparities in expertise and resources.
Keywords
automation, contouring, quality assurance, medical physics, peer review