Dissertation
Advancing online adaptive radiotherapy: motion management and MR-guided planning using deep reinforcement learning
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Spring 2026
DOI: 10.25820/etd.008421
Abstract
Adaptive radiotherapy (ART) on MRI-guided linear accelerators (MR-Linacs) offers the potential to improve treatment precision by adapting radiation plans to daily anatomical changes. However, realizing this potential is limited by challenges in real-time motion monitoring and the time-intensive nature of online treatment plan optimization. This dissertation presents a series of interconnected advances addressing these limitations on the Elekta Unity 1.5T MR-Linac.
First, a T1-weighted cine MRI sequence was developed and validated for real-time respiratory motion management. Phantom experiments demonstrated sufficient spatial resolution, signal-to-noise ratio, and temporal resolution for motion tracking. In-vivo evaluation across 11 patients showed improved contrast-to-noise ratio and tracking success relative to the standard balanced steady-state free precession (bSSFP) cine sequence, particularly for liver tumors where T2/T1-weighted contrast is insufficient for reliable target visualization.
Second, surrogate gating strategies were developed and evaluated for the Elekta Unity comprehensive motion management (CMM) system. These strategies enable respiratory-gated beam delivery for treatment sites where the target cannot be directly visualized in cine imaging, including central lung, pancreas, liver, and prostate targets. This work is especially important as the developed T1-weighted cine MRI is not yet clinically available.
Third, as a step towards accelerating online treatment planning, a rapid dose calculation algorithm was developed for repeated intensity-modulated radiation therapy (IMRT) dose computation. The algorithm precomputes correction beamlets and patient-specific open field doses, enabling full IMRT dose calculations with high accuracy as validated against a clinical treatment planning system across prostate and lung cases, including scenarios with metallic implants and low-density tissue inhomogeneities.
Fourth, a tandem reinforcement learning (RL) framework was developed for end-to-end volumetric modulated arc therapy (VMAT) machine parameter optimization (MPO) for localized prostate cancer. Using proximal policy optimization, tandem networks were trained to predict multi-leaf collimator positions and monitor unit values directly, bypassing the need for conventional inverse optimization. The framework generated completely physical and clinically acceptable treatment plans in seconds. All plans met predefined dose-volume constraints and demonstrated improved organ-at-risk sparing compared to reference plans.
Fifth, a variational autoencoder (VAE) was developed for state representation learning to support generalization of the RL framework to more complex treatment sites. The VAE learned compact latent representations of beam’s-eye-view projections encoding aperture, target, and organ-at-risk information, demonstrating high reconstruction fidelity and a structured latent space with smooth interpolation properties suitable for reinforcement learning.
Together, these contributions address critical bottlenecks in real-time motion management and online plan adaptation, advancing the feasibility of efficient, accurate, and accessible MRI-guided adaptive radiotherapy.
Details
- Title: Subtitle
- Advancing online adaptive radiotherapy: motion management and MR-guided planning using deep reinforcement learning
- Creators
- Nathan Shaffer
- Contributors
- Joel J. St-Aubin (Advisor)Joseph M. Reinhardt (Committee Member)Blake R. Smith (Committee Member)Sajan G. Lingala (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Biomedical Engineering
- Date degree season
- Spring 2026
- DOI
- 10.25820/etd.008421
- Publisher
- University of Iowa
- Number of pages
- xvii, 193 pages
- Copyright
- Copyright 2026 Nathan Shaffer
- Language
- English
- Date submitted
- 04/22/2026
- Description illustrations
- Illustrations, graphs, charts, tables
- Description bibliographic
- Includes bibliographical references (pages 174-190).
- Public Abstract (ETD)
Radiation therapy is a common cancer treatment that uses focused beams of radiation to destroy tumors. One of the biggest challenges is ensuring the radiation hits the tumor precisely while avoiding healthy tissue. Tumors can shift during treatment due to breathing, digestion, and other natural body movements, and a patient’s anatomy can change over the days or weeks of a treatment course. Adaptive radiotherapy addresses this by adapting the radiation treatment plan to match the patient’s current anatomy at each treatment, but the speed of these adaptations has remained a significant challenge.
This dissertation presents new tools to make adaptive radiotherapy faster and more accurate on an MRI-guided treatment machine, which combines a radiation beam with an MRI scanner so clinicians can image the body during treatment. First, we developed a new type of real-time MRI scan that provides better image contrast for certain tumors, particularly in the liver, making it easier for the system to track tumor motion during treatment. We also developed strategies to track tumors that are difficult to see directly by monitoring nearby anatomical landmarks as surrogates.
To speed up the treatment planning process, we developed a fast algorithm for calculating radiation dose and then used it to train an artificial intelligence system based on reinforcement learning, the same AI training technique famously used to play Go and fine-tune large-language models. Instead of requiring a human planner to manually adjust hundreds of parameters over many minutes with an optimizer, our AI system learns to generate treatment plans in seconds by repeatedly practicing and improving through trial and error. Finally, we developed the first stage of a new AI network that learns to efficiently represent the complex treatment geometry, with the goal of extending this approach to more challenging cancer sites.
Together, these advances work to accelerate and simplify the adaptive radiotherapy process, reducing treatment times, improving accuracy, and ultimately improving adaptive therapy access for cancer patients.
- Academic Unit
- Roy J. Carver Department of Biomedical Engineering
- Record Identifier
- 9985177272902771
Metrics
1 Record Views