Synthesizing High-Resolution 4D Cine MR Images for Mobile Tumor Motion Assessment
J Lee1,2*, H Gach1, E Laugeman1, 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) Oncofost Inc., Seoul, KR
Presentations
SU-H300-IePD-F9-5 (Sunday, 7/10/2022) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 9
Purpose: CBCT-guided adaptive radiotherapy (CT-ART) with optical surface imaging has been introduced for mobile tumors in abdominothoracic regions (Ethos-Identify). However, the correlation between the motions of the skin surface and internal tumors has not been fully assessed. Real-time cine MRI is currently functional but limited to the 2D motion of the tumors. In this study, we synthesize high-resolution 4D cine MR images by using intentional overfitting of the deep learning model to assess 3D internal motion corresponding to the skin surface motion.
Methods: The deep learning-based super-resolution network was trained in two-fold: 1) 'basic' training using 24 pairs of 0.35T 3D abdominal MR scan images from the eight individuals, and 2) 'intentional overfit' training using 3D breath-hold MR images of three volunteers. Afterward, the trained model was applied to the same applicant's low-resolution free-breathing cine 3D MR images to generate high-definition images. Since the image features were different, the model was trained and applied separately for each plane. The evaluation of the results was conducted using metrics that no reference is required.
Results: For the three volunteers, each plane’s images of low-resolution free-breathing 2D cine MR were converted into high-resolution. By arranging these in chronological order, the movement due to breathing in each plane was possible to be observed. In addition, the results of quantifying the difference in quality between the LR image and the proposed SR images by using an index that can evaluate the spatial quality (BRISQUE), naturalness (NIQE), and perception-based index (PIQE) without a reference image were as follows. 1) BRISQUE: 48.97±1.98→45.87±1.76 (axial), 39.90±1.08→36.21±0.72 (sagittal), 43.05±0.69→36.21±0.72 (coronal), 2) NIQE: 11.38±0.99→4.55±0.42, 15.14±0.59→11.86±0.55, 16.18±1.00→11.86±0.55, 3) PIQE: 58.08±2.95→45.01±4.03, 62.08±3.80→41.75±4.18, 60.83±3.56→41.75±4.18.
Conclusion: Through this study, we discovered the possibility of tracking tumor movement and quantifying it by acquiring high-resolution images in real-time according to the patient's breathing in 3D cine MR.
Keywords
Image-guided Therapy, Interventional MRI, High-resolution Imaging
Taxonomy
Contact Email