S Shan^1,2, Y Gao³, P Liu^1,2, T Reynolds¹*, B Dong², H Sun³, M Li³, G Liney², F Liu³, P Keall^1,2, D Waddington^1,2, (1) ACRF ImageX Institute, University of Sydney, Sydney, NSW 2015, Australia (2) Ingham Institute for Applied Medical Research, Liverpool, NSW 2170, Australia(3)School of Information Technology & Electrical Engineering, University of Queensland, Brisbane, QLD 4067, Australia

• 
  T Reynolds

TH-F-BRC-3 (Thursday, 7/14/2022) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Ballroom C

Purpose: Superior soft-tissue contrast on MRI-Linacs enables the application of real-time image guidance and adaptive tumor tracking for radiotherapy treatments. However, respiratory motion can lead to artifacts that reduce MR image quality and affect the accuracy of tumor tracking. Here, we investigate the use of an interpretable model-driven deep-learning unrolling network to facilitate rapid reconstruction of motion-corrected thorax images and respiratory signal extraction for real-time tumor tracking on an MRI-Linac.

Methods: An unrolling network was trained to prospectively reconstruct dynamic low-resolution motion-free images. The images were used to extract respiratory signals for tumor tracking during treatment. After treatment, high-resolution images were reconstructed from the network with retrospective self-gating for intrafraction treatment verification. The unrolling network was implemented in Pytorch with an architecture consisting of four blocks, which starts with a data fidelity calculation, followed by nonlinear transforms with a soft-thresholding operation. Datasets of 3000 training thorax images and 450 validation images from ten volunteers were encoded with a golden-angle radial acquisition to generate raw k-space data. Our deep-learning-reconstruction approach was validated with a digital CT/MRI breathing XCAT (CoMBAT) phantom and experimentally with a motion phantom and a healthy volunteer on an MRI-Linac.

Results: Digital phantom results showed that the unrolling network successfully estimated respiratory signals and reconstructed motion-corrected images with comparable error metrics (RMSE=0.02, SSIM=0.93) to the least square (LSQR) algorithm (RMSE=0.02, SSIM=0.92), a benchmark for radial reconstruction whilst being fast enough for real-time implementation (50ms vs 2s). Experimental motion phantom and volunteer tests showed that retrospective self-gated image reconstruction resulted in sharper edges and fewer motion artifacts when compared to conventional LSQR reconstruction.

Conclusion: We have demonstrated rapid motion-corrected image reconstruction and real-time respiration estimation from an MRI-Linac with an unrolling network. Experimental implementation promises to reduce motion artifacts and reconstruction latency for dynamic tumor tracking.

Funding Support, Disclosures, and Conflict of Interest: David Waddington and Paul Liu are supported by a Cancer Institute NSW Early Career Fellowship. This work has been funded by the Australian National Health and Medical Research Council Program Grant APP1132471.

Keywords

MRI

Taxonomy

IM- MRI : Machine learning, computer vision

Contact Email