Click here to

  1. Home
  2. Meeting Program
  3. All Sessions
  4. Abstracts
Session: MR Guidance for Motion Management [Return to Session]

Real-Time Target Tracking Enhancement in 0.35T MRgRT Using Super-Resolution Reconstruction

J Lee1,2, Z Ji1, S Marasini1, J Chun2,3*, J Kim2,3, T Kim1, (1) Washington University School of Medicine in St. Louis, St. Louis, MO, (2) Yonsei University College of Medicine, Seoul, KR, (3) Oncosoft Inc., Seoul, KR

Presentations

WE-C1000-IePD-F2-5 (Wednesday, 7/13/2022) 10:00 AM - 10:30 AM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 2

Purpose: Image quality is a key component in MR-guided radiotherapy (MRgRT). In 0.35T MRgRT, fast 2D cine MRI is utilized for real-time imaging, while target tracking accuracy challenges due to low-resolution image quality. In this study, by using the deep learning technique, real-time 2D cine images were converted to super-resolution images (SR), and the clinical feasibility of SR was tested using target tracking evaluation.

Methods: A deep learning network for super-resolution was first trained using 1,855 pairs of low- and high-resolution abdominal MR images, then the network was customized with sagittal pair of MR images of a patient prepared for the test. Super-resolution images were obtained by applying the trained model to real-time LR 2D cine MRI acquired while a test patient was breathing. After uploading LR and SR images to the ViewRay Simulator, we checked whether the target tumor could be tracked. Results were quantified by using the image evaluators without reference images to evaluate the spatial quality (BRISQUE), naturalness (NIQE), and perception-based quality (PIQE).

Results: From ViewRay Simulator, target tracking was performed with each frame of original LR (before and after interpolated by the system), and proposed SR. As the resolution increased, the tracking target was getting distinct, indicated by the thin contour in the proposed SR compared to that of the original LR. That clinically meant that uncertainty of the tracking tacking can be improved using the proposed technique. Meanwhile, the mean±std for each image were as follows in order: 1) BRISQUE: 32.55±3.47 (LR), 47.39±0.79 (interpolated LR), 20.34±1.49 (SR) 2) NIQE: 18.88±0.00, 5.96±0.27, 3.94±0.26 3) PIQE: 55.86±2.48, 74.31±2.49, 42.04±2.46.

Conclusion: Through this study, we demonstrated SR image generation from 0.35T real-time LR cine MRI and the feasibility of the target tracking on the proposed SR images for MRgRT.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email

Share:

Additional Links