Implementation of Super-Resolution Imaging On An MRI-Linac
J Grover1, P Liu1,2, B Dong2, S Shan1,2, B Whelan1,2, P Keall1,2*, D Waddington1,2, (1) ACRF Image X Institute, Faculty of Medicine and Health, The University of Sydney, NSW, AU, (2) Department of Medical Physics, Ingham Institute for Applied Medical Research, NSW, AU.
Presentations
TU-J430-BReP-F2-5 (Tuesday, 7/12/2022) 4:30 PM - 5:30 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 2
Purpose: Real-time MRI is traditionally constrained to low-resolution acquisitions to reduce imaging latency. Low-latency, high-resolution images could improve the accuracy of real-time tumor-tracking. Here, we present an implementation of super-resolution on an MRI-Linac to enable real-time high spatiotemporal resolution imaging.
Methods: We implemented an Enhanced Deep residual network for single image Super-Resolution (EDSR) that increases image resolution by 4×. The pre-trained model was fine-tuned on 136 low/high-resolution MR image pairs acquired on our MRI-Linac. Super-resolution methods were deployed in real-time on our MRI-Linac using Gadgetron. We tested our implementation by imaging the brain at a super-resolution of 256×256 (input acquired at 64×64) and comparing to a ground truth conventional 256×256 acquisition using a normalized root-mean-square-error (NRMSE). Super-resolution accuracy was compared to interpolation by zero-filling in k-space as the baseline method. Free-breathing cine-MRI scans of the thorax were acquired at a super-resolution of 256×256 (64×64 input) to demonstrate the feasibility of using super-resolution methods for tumor tracking.
Results: Super-resolution was successfully implemented on the MRI-Linac. EDSR and interpolation methods upsampled a 64×64 input image to 256×256 in less than 6ms on a NVIDIA Tesla P100 GPU. Acquisition time for a 256×256 super-resolution image was 6 times faster than a conventional 256×256 acquisition (approximately 600ms vs 3600ms). For brain imaging, the EDSR super-resolution images had an error (NRMSE 7.82%) when compared to the ground-truth, which is a lower error than found when upsampling by interpolation through zero-filling of k-space (NRMSE 8.98%). For thorax imaging, super-resolution images enhanced the images, producing clearer anatomic boundaries suitable for real-time tumor tracking.
Conclusion: A deep learning-based super-resolution methodology was implemented on an MRI-Linac, exhibiting lower error than conventional upsampling by interpolation through zero-filling in k-space. Integration of these techniques with real-time adaptive radiotherapy is expected to improve MR-guided tumor tracking accuracy.
Funding Support, Disclosures, and Conflict of Interest: David Waddington is supported by a Cancer Institute NSW Early Career Fellowship 2019/ECF1015. This work has been funded by the Australian National Health and Medical Research Council Program Grant APP1132471.
Keywords
Taxonomy
IM- MRI : Machine learning, computer vision
Contact Email