Click here to

Session: Deep Learning Image Formation and Motion Management [Return to Session]

Improving Tumor Tracking Accuracy During MR-Guided Radiotherapy Using Pre-Trained Super-Resolution Neural Networks

P Liu1,2*, D Waddington1,2, J Grover1, P Keall1,2 (1) ACRF Image X Institute, Faculty of Medicine and Health, The University of Sydney, NSW, AU (2) Department of Medical Physics, Ingham Institute of Applied Medical Research, NSW, AU


WE-C930-IePD-F5-2 (Wednesday, 7/13/2022) 9:30 AM - 10:00 AM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 5

Purpose: During MR-guided radiotherapy, real-time imaging is acquired at low resolutions to minimize latency. This degrades the geometric accuracy of target tracking. Pre-trained super-resolution neural networks can up-sample images to higher resolutions whilst preserving detail. The purpose of this work is to demonstrate the use of super-resolution to improve the accuracy of real-time target tracking.

Methods: Target tracking performance was analyzed on 21,000 lung cineMRIs from 14 radiotherapy patients. Images were down-sampled in k-space from 256x256 pixel resolution to 128x128 and 64x64 resolutions. The down-sampled images were then up-sampled at 2x and 4x back to 256x256 pixels. Up-sampling was performed with three real-time super-resolution networks (ESPCN, FSRCNN, LapSRN) and compared to baseline methods (no interpolation, bicubic interpolation).For each up-sampling method, the radiotherapy target centroid position was calculated using template matching with a semi-manually contoured target template. Accuracy was quantified by comparing the calculated positions to the ground truth positions calculated using template matching on the original images.

Results: FSRCNN was the best performing method overall. At 2x up-sampling, tracking using FSRCNN reduced the mean geometric error by 70.1% compared to no interpolation and 43.1% compared to bicubic interpolation. At 4x up-sampling, the mean error was reduced by 37.6% and 13.2%, respectively. The error reduction varied by patient, particularly at 4x up-sampling (-9% to 32% compared to bicubic) but was more consistent at 2x (17% to 62%).At 2x up-sampling, FSRCNN provided a more accurate target position than bicubic interpolation in 20.3% of images and less accurate in 4.3% of images, with the remainder giving the same position. At 4x, FSRCNN performed better in 22.7% of images and worse in 11.2%.

Conclusion: Pre-trained super-resolution networks are freely available in the open-domain. They are fast and can be easily implemented within the target tracking workflow to improve tracking accuracy.




IM/TH- MRI in Radiation Therapy: MRI/Linear accelerator combined- IGRT and tracking

Contact Email