Dissertation
Imago Restituta: synthetic generation and deep learning detection/correction of respiratory motion artifacts in 4DCT
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Spring 2026
DOI: 10.25820/etd.008524
Abstract
Radiation therapy planning for lung cancer relies heavily on Four-Dimensional Computed Tomography (4DCT) to accurately characterize respiratory motion and target malignant tumors. However, irregular patient breathing during acquisition frequently leads to motion artifacts, such as interpolation, duplication, and truncation artifacts, which degrade image fidelity and distort anatomical geometry. These artifacts introduce significant uncertainty into clinical workflows, complicating tumor delineation and potentially confounding advanced treatment strategies like functional avoidance radiation therapy. While advancements in scanner technology and analytical correction methods have been introduced, they often suffer from high computational costs or limited availability, preventing widespread clinical integration. Furthermore, the development of automated solutions has historically been constrained by the scarcity of paired ground-truth data, as obtaining both artifact-laden and artifact-free images of the same respiratory phase is clinically impractical.
Our work attempts to solve this problem by introducing a comprehensive deep learning framework which encompasses the synthesis, detection, and correction of 4DCT artifacts. We hypothesized that by synthetically modeling the morphological characteristics of motion artifacts, we could generate the large-scale datasets necessary to train robust neural networks for automated quality assurance. The first part of this work investigated the creation of various synthetic artifacts—including interpolation, duplication, truncation, and phase shadow—to understand their morphology. We subsequently utilized real artifact data, augmented specifically with synthetic interpolation artifacts, to train AD2, a 2D U-Net model. While implemented as a single-class segmentation network, we demonstrated that this model could accurately localize diverse artifact types simultaneously, achieving performance metrics that matched human observer sensitivity and outperformed traditional analytical detection algorithms.
The second body of this work concerns the development of the Multi-Phase Artifact Reduction/Correction (MARC) framework, a 4D generative adversarial model designed to restore anatomical integrity in corrupted scans. To achieve this, we proposed a novel image-to-image translation approach that leverages the temporal redundancy of 4DCT by using adjacent respiratory phases as anatomical context for reconstruction. We demonstrated that MARC could correct interpolation artifacts with fidelity comparable to state-of-the-art registration-based methods while reducing computational processing time from hours to seconds. Furthermore, our evaluation showed that this approach successfully preserved non-artifactual anatomy better than whole-lung regression techniques and generalized effectively to real-world clinical data.
Details
- Title: Subtitle
- Imago Restituta: synthetic generation and deep learning detection/correction of respiratory motion artifacts in 4DCT
- Creators
- Joshua W. Carrizales
- Contributors
- Joseph M. Reinhardt (Advisor)Gary E. Christensen (Committee Member)Sarah E. Gerard (Committee Member)John E. Bayouth (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.008524
- Publisher
- University of Iowa
- Number of pages
- xiv, 97 pages
- Copyright
- Copyright 2026 Joshua W. Carrizales
- Language
- English
- Date submitted
- 03/11/2026
- Description illustrations
- Illustrations, graphs, charts, tables
- Description bibliographic
- Includes bibliographical references (pages 84-97).
- Public Abstract (ETD)
Lung cancer is a leading cause of cancer-related deaths, and radiation therapy is a primary way to treat it. To plan this treatment, doctors use a technology called Four-Dimensional Computed Tomography (4DCT), which creates a movie of the patient’s lungs to track breathing motion. However, patients often breathe irregularly during these scans, causing errors in the images known as ”artifacts”. These artifacts are similar to the blurs or distortions seen in a panoramic photo when a person moves; they can hide the tumor or distort the location of healthy organs, making treatment planning difficult.
My research focuses on restoring the image to fix these errors using Artificial Intelligence (AI). I developed a three-part process to ensure doctors have accurate images. First, I created realistic simulations of these errors to teach computers what to look for. Second, I trained AI models to automatically detect and locate these glitches within the scans. Finally, I built specialized AI models to correct the artifacts, effectively erasing the errors and filling in the correct details.
By automating the detection and correction of these image errors, this work helps ensure that radiation therapy is targeted precisely at the tumor while sparing healthy tissue. This technology provides a reliable way to improve image quality for cancer patients, ensuring that doctors base their decisions on the true anatomy rather than distorted images.
- Academic Unit
- Roy J. Carver Department of Biomedical Engineering
- Record Identifier
- 9985176976402771
Metrics
1 File views/ downloads
1 Record Views