Respiratory motion augmentation for personalized super-resolution (RMApSR) of 3D cine MR images in MRI-guided radiotherapy.
Authors
Affiliations (5)
Affiliations (5)
- Medical Physics and Biomedical Engineering Lab (MPBEL), Yonsei University College of Medicine, Seoul, Republic of Korea; Department of Radiation Oncology, Yonsei Cancer Center, Heavy Ion Therapy Research Institute, Yonsei University College of Medicine, Seoul, Republic of Korea.
- Department of Radiation Oncology, Gangnam Severance Hospital, Yonsei University College of Medicine, Seoul, Republic of Korea.
- Medical Physics and Biomedical Engineering Lab (MPBEL), Yonsei University College of Medicine, Seoul, Republic of Korea; Department of Radiation Oncology, Yonsei Cancer Center, Heavy Ion Therapy Research Institute, Yonsei University College of Medicine, Seoul, Republic of Korea; Oncosoft Inc., Seoul, Republic of Korea.
- Department of Radiation Oncology, Gangnam Severance Hospital, Yonsei University College of Medicine, Seoul, Republic of Korea. Electronic address: [email protected].
- Department of Radiation Oncology, Washington University in St. Louis, St Louis, MO, United States of America.
Abstract
The technical implementation of volumetric magnetic resonance imaging (MRI) has the potential to substantially advance tumor tracking in MRI-guided radiotherapy (MRIgRT). However, existing three-dimensional cine MRI techniques remain constrained by tradeoffs between spatial and temporal resolution, with spatial resolution often limited to a voxel spacing of 5-6 mm. In this study, we proposed a respiratory motion augmentation (RMA) method for personalized super-resolution (pSR) reconstruction to achieve sufficiently high spatial and temporal resolutions for real-time volumetric MRI. The proposed method enhances the motion robustness of pSR networks by synthesizing high-resolution MR images across multiple respiratory phases using deformable image registration (DIR). A conventional SR network was initially trained on a public dataset of 78 patients and subsequently personalized using 1.5-T MR-LINAC data from 12 cancer patients (10 abdominal and 2 prostate cancer cases). For network personalization, the proposed RMA approach was applied by deforming a breath-hold high-resolution image to match free-breathing low-resolution images using DIR, thereby generating multiple paired datasets for each patient. The RMApSR network substantially improved both image quality and segmentation accuracy. The network achieved increases of 1.7% in the peak signal-to-noise ratio, 6.2% in the structural similarity index measure, and 10.7% in the Dice similarity coefficient compared with those of the conventional pSR while maintaining consistent performance across respiratory phases. Personalized training, which requires 13.2 min, can be performed in parallel with standard MRIgRT procedures. Combined with an inference time of 72 ms per volume, our approach demonstrates high clinical feasibility. This motion-aware personalization strategy represents a significant advancement toward achieving high spatiotemporal resolution requirements for volumetric MRIgRT implementation.