Epigastric CBCT Motion Artifacts Correction Based on Four-Dimensional Motion Simulation and Swin-Transformer-Based Generative Adversarial Network
Abstract
Summary
Background: Respiratory and gastrointestinal movements induce motion artifacts in cone-beam computed tomography (CBCT), compromising image quality for adaptive radiotherapy. Existing approaches, including phase-binned four-dimensional (4D) imaging and deep-learning based correction methods, are hindered by noise, computational inefficiency, and reliance on scarce artifact-free reference data, limiting their clinical utility in precise tumor localization.
Purpose
This study aims to enhance the clinical utility of CBCT in adaptive radiotherapy by correcting the motion artifacts caused by respiratory and gastrointestinal movements in upper abdominal imaging. The proposed framework integrates a 4D motion simulation with a Swin-Transformer-based generative adversarial network (SWTGAN) to enhance image quality for precise tumor localization.
Methods
We developed a novel SWTGAN framework that integrates a Swin-Transformer generator and a PatchGAN discriminator. Motion artifacts were simulated using 4D computed tomography (CT) data to generate pseudo-CBCT images, which were paired with high-quality planning CT scans for training. The generator captured global motion patterns and local anatomical details, whereas adversarial training optimized artifact suppression. Clinical datasets from 25 patients were used for evaluation.
Results
SWTGAN outperformed traditional methods, achieving the lowest root mean square error values of 52.60 HU and 33.50 HU in the thoracic and abdominal regions, respectively. It also achieved the highest structural similarity index measure scores of 0.9769 and 0.9758, as well as optimal spatial uniformity with SNU values of 1.32% and 1.07%, closely matching the ground-truth metrics. The visual assessment confirmed effective artifact reduction in critical anatomical regions.
Conclusion
The SWTGAN effectively corrects motion artifacts through the fusion of global and local features and adversarial training, thereby reducing reliance on scarce artifact-free clinical data. This framework makes CBCT a reliable tool in adaptive radiotherapy, bridging the gap between dynamic imaging and precision treatment.