2291 - Respiration Augmented Personalized Super Resolution for Real-Time 3D MR-Guided Adaptive Radiotherapy
Presenter(s)
Y. Yoon1,2, J. Sung3, J. W. Kim3, J. S. Kim1,4, T. Kim5, and J. Kim6; 1Department of Radiation Oncology, Yonsei Cancer Center, Heavy Ion Therapy Research Institute, Yonsei University College of Medicine, Seoul, Korea, Republic of (South), 2Department of Radiation Oncology, Washington University School of Medicine, St. Louis, MO, 3Department of Radiation Oncology, Gangnam Severance Hospital, Yonsei University College of Medicine, Seoul, Korea, Republic of (South), 4Oncosoft Inc., Seoul, Korea, Republic of (South), 5Wash U School of Medicine, Department of Radiation Oncology, St. Louis, MO, 6Department of Radiation Oncology, Gangnam Severance Hospital, Yonsei University College of Medicine, Seoul, Republic of Korea, Seoul, Korea, Republic of (South)
Purpose/Objective(s): Deep learning–based super-resolution (SR) network can enhance the spatial resolution of medical imaging, including magnetic resonance imaging (MRI). However, developing SR networks for the abdomen region is challenging since high-resolution (HR) MR images can only be acquired at a specific respiratory phase while acquiring multi-phase real-time scans is feasible in low-resolution (LR). We propose a respiration-augmented personalized SR (RApSR) network tailored for real-time volumetric MRI-guided radiation therapy (MRIgRT).
Materials/Methods: We collected two MRI datasets for network training. A public data set of 78 patients was used to create a generalized SR (gSR) network; LR MRI scans were generated by performing k-space down-sampling with a factor of four on HR MR images (mean image dimension: 80 × 80 × 36, mean spacing: 1.2 × 0.9 × 3.7 mm3). Another data set was collected for 10 abdominal cancer patients who underwent MRIgRT using 1.5-T MR-Linac, each of which includes a breath-hold HR scan (image dimension: 320 × 320 × 136, spacing: 1.25 × 1.25 × 1.5 mm3) and 100 free-breathing LR scans (image dimension: 80 × 80 × 34, spacing: 5.0 × 5.0 × 6.0 mm3, temporal resolution: ). To compensate for phase mismatch between HR and LR MRI scans, data augmentation was performed by deforming a breath-hold HR image to multi-phase LR images using deformable image registration, resulting in six HR/LR pairs for each patient. The gSR network was trained with the augmented data to develop the RApSR network. For image quality evaluation, we compared SR MRI scans with phase-matched breath-hold MRI scans, only for the same respiration phase, utilizing peak signal-noise ratio (PSNR) and structural similarity measure index (SSIM). The performance of the proposed network was further explored by calculating Dice similarity coefficient (DSC) of 25 organs segmented using a publicly available automatic segmentation software (TotalSegmentator).
Results: The RApSR network significantly improved both image quality and segmentation accuracy. Compared to linear interpolation, gSR, and personalization without respiratory augmentation, it achieved increases of up to 5.8% in PSNR (33.57±0.80 vs. 31.73±0.61, 32.10±0.65, and 33.01±0.74), 41.0% in SSIM (0.86±0.03 vs. 0.61±0.06, 0.70±0.05, and 0.80±0.04), and 36.1% in DSC (0.83±0.12 vs. 0.61±0.20, 0.64±0.18, and 0.74±0.15). Additionally, the average SR reconstruction time was 6.5±0.2 ms per 3D free-breathing LR MRI volume.
Conclusion: The proposed RApSR reconstruction significantly enhanced image and segmentation quality of 3D free-breathing LR MR images, achieving sufficiently high spatial resolution and temporal resolution for real-time volumetric MR imaging. Therefore, this approach represents a valuable technological advancement for clinical implementation of real-time volumetric MRIgRT.