Purpose: Spatiotemporal resolution of MR images is limited by hardware and acquisition parameters, especially at low magnetic fields where long imaging times are required to improve SNR and feature identification. These factors are vital to target tracking and beam gating during MR-guided radiotherapy (MRgRT). Super resolution (SR) reconstruction with deep-learning frameworks provides the ability to generate SR images from low resolution (LR) images. This study applies the SR framework to 2D cine MRI acquired on an MRgRT system during pancreatic cancer treatment.
Methods: Real time 4 frames-per-second (FPS) cine MR images were collected retrospectively for fifty pancreatic cancer patients treated on the ViewRay MR-linac system. A previously trained super resolution framework was used to generate SR images from the original 3.5x3.5mm² pixel images, with a new resolution of 0.9x0.9mm². SR images were compared to the original LR using inherent metrics including: NIQE, PIQE, and BRISQUE. Inherent image metrics score from 0 to 100, 0 indicates an ideal image. SR images were also compared to paired high resolution (HR) 3D images using SSIM and anterior wall intensity gradient as a surrogate for edge sharpness.
Results: SR images had statistically significant improvements for inherent metrics, compared to the LR images. SR images scored a mean (±SD) of 3.17±0.18 for NIQE, 34.58±3.38 for PIQE, and 29.65±2.98 for BRISQUE. LR images scored a mean (±SD) of 7.40±0.69 for NIQE, and 69.46±1.52 for PIQE, and 42.48±0.98 for BRISQUE. SR images also had greater SSIM values and sharper line profiles compared to LR images with SSIM of 0.633±0.063 and 0.587±0.067, and gradients of 17.71±3.17 and 14.39±3.51, for SR and LR images, respectively.
Conclusion: Super resolution images provided improved image quality and target identification without introducing artifacts or anatomical inconsistencies. This leads the way for real time SR image generation during MRgRT for target tracking and gating.