DOI
https://doi.org/10.25772/T6Y1-A062
Defense Date
2025
Document Type
Thesis
Degree Name
Doctor of Philosophy
Department
Radiation Oncology
First Advisor
William Song
Second Advisor
Siyong Kim
Third Advisor
Lulin Yuan
Fourth Advisor
Monica Ghita
Abstract
Brachytherapy (BT) is a critical treatment modality for cervical cancer that delivers ionizing radiation directly to the tumor using intracavitary (IC) and/or interstitial (IS) applicators. In high dose-rate (HDR) BT, a remote afterloader guides the radioactive source to specific dwell positions for defined durations, known as dwell times, allowing for a conformal dose to the tumor while minimizing exposure to surrounding organs at risk (OAR). However, due to the inherent nature of ionizing radiation, some dose to OARs is unavoidable, creating a fundamentally conflicting objective in treatment planning: maximizing tumor coverage while minimizing OAR dose. This characterizes the problem as a multi-criteria optimization (MCO) task, where multiple Pareto-optimal solutions are required to represent clinically relevant trade-offs. The nonconvex nature of optimization based on the dose-volume histogram (DVH) further complicates the process, making it computationally intensive and impractical for routine clinical implementation. In addition, time efficiency is critical, as patients typically remain under anesthesia or experience discomfort during applicator placement and while awaiting finalized treatment plans. To address these challenges, this thesis proposes an automated treatment planning framework that integrates deep learning-based dose prediction with MCO to enable faster and more robust optimization, thus improving clinical workflow, improving plan quality, and supporting high-quality patient-centered care. Current commercial planning algorithms in HDR BT often rely on manual fine-tuning of objective functions and/or dwell times. Inverse planning algorithms typically generate only one treatment plan per optimization run, without guaranteeing the fulfillment of clinical goals in the first attempt. Consequently, planning becomes an iterative and time-consuming process, and the quality of the plan is highly dependent on the experience and skill of the planner. To overcome this limitation, we introduce a knowledge-based planning model (KBP) that predicts D2cc values, detects suboptimal plans, and improves plan quality - particularly relevant for direction-modulated brachytherapy (DMBT), a new applicator technology with limited clinical experience. The KBP-based plan serves as the foundation for a deep learning-based dose distribution prediction model, which is incorporated into the automated framework. This approach facilitates the refinement of suboptimal plans and allows personalized quality control, offering a reliable and accurate tool for independent plan evaluation. Although deep learning-based dose prediction provides a strong starting point, the generated distributions may not be fully optimal. Furthermore, the iterative one plan- per-optimization strategy used in current clinical workflows restricts the ability to explore trade-offs between target coverage and OAR sparing. A significant gap also exists in optimization techniques capable of efficiently handling complex applicators such as DMBT, where determining optimal dwell positions and depths is especially challenging due to the large number of degrees of freedom and variability in patient anatomy. To address these issues, this work integrates a novel parallelized CPU-based MCO algorithm into the planning pipeline. This algorithm enables rapid and automated exploration of the Pareto surface, improving decision-making efficiency and overall treatment quality. The implemented CPU-based MCO algorithm is benchmarked against the standard inverse planning algorithms currently used in the clinic. It features a novel parallel optimization scheme capable of generating thousands of Pareto-optimal plans within seconds. Plan quality is evaluated in comparison to plans produced by the clinical benchmark (BVTPS), and the time required for MCO plan generation is recorded. By integrating the KBP-guided base plan, deep learning-based dose prediction, and the MCO engine, a cohesive automated treatment planning workflow is established. The eight chapters of this thesis present the methods, findings, and scientific contributions across both conventional and advanced DMBT applicators, culminating in a recommended framework for clinical automation of BT planning.
Rights
© The Author
Is Part Of
VCU University Archives
Is Part Of
VCU Theses and Dissertations
Date of Submission
7-27-2025