Frequency Schrödinger bridge diffusion model for fast MRI reconstruction in abdominal MR-guided online adaptive radiotherapy.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiation Oncology, Chinese Academy of Medical Sciences, Peking Union Medical College, Beijing, China.
- School of Biomedical Engineering, Tsinghua University, Beijing, China.
- MedMind Technology Co., Ltd, Beijing, China.
- School of Biomedical Engineering and the Institute of Medical Robotics, Shanghai Jiao Tong University, Shanghai, China.
Abstract
Online adaptive radiotherapy (ART) employs modalities like CBCT, CT, and MRI. For abdominal malignancies, MRI stands out with superior soft-tissue contrast, no ionizing radiation, and functional imaging capabilities. Yet, MRI-guided ART for such cases is challenged by long treatment time. Thus, boosting efficiency, especially via accelerated MRI scanning, is vital. This study presents the first accurate and fast MRI reconstruction method tailored for the Unity MR-Linear system. By slashing MRI acquisition time, it aims to enhance patient compliance, reduce intra-fraction anatomical motion, and optimize clinical resource use. A frequency Schrödinger bridge diffusion (FSBD) model for MRI reconstruction (converting 22-second MRI to full-duration MRI) was proposed, based on a retrospective analysis of 31 fractional MRI scans from 14 patients. Its reconstruction performance was compared with state-of-the-art methods: paired diffusion-based Refusion, and unpaired CycleGAN and CSGAN. Of the initially acquired 40 fractional scans, 9 examinations were excluded following predefined exclusion criteria, leaving 31 valid cases. Dataset was split strictly at patient level: 23 fractional scans for training and the remaining independent patients for testing. An extra baseline using pairwise full-duration MRI from different fractions of identical patients was supplemented for reliability verification. All metrics were calculated on complete 3D volumetric MRI rather than individual 2D slices. Quantitative evaluation shows the FSBD model's superiority. It achieved significantly lower error metrics: MAE (86.00) and RMSE (123.42), outperforming Refusion (MAE: 99.95, RMSE: 172.55), CycleGAN (MAE: 112.92, RMSE: 149.16), and CSGAN (MAE: 103.04, RMSE: 162.29). In image quality, FSBD's PSNR (28.44 dB) and SSIM (0.885) surpassed Refusion (26.77 dB / 0.870), CycleGAN (25.82 dB / 0.881), and CSGAN (26.34 dB / 0.878). Moreover, it delivered high-fidelity reconstruction with an efficient 20-s inference time. This study is the first to propose a sub-one-minute MRI imaging method for the Unity MR linear accelerator in an MR-guided online adaptive radiotherapy (ART) workflow. Experimental results confirm the method's significantly better reconstruction performance than state-of-the-art approaches. Total MRI imaging time dropped drastically from 3 min (full-duration) to 42 s (22 seconds for MR imaging plus 20 seconds for reconstruction). This reduction brings notable clinical benefits, including improved patient compliance and fewer intra-fractional anatomical variations.