Real-Time Volumetric MRI for MR-Guided Abdominal Radiotherapy Using a Cyclegan Model
Abstract
Purpose
The excellent soft-tissue contrast provided by MRI offers improved 3D tumour delineation for radiotherapy beam adaptation on MRI-Linacs. However, long acquisition and reconstruction times pose a barrier to achieving volumetric MRI in real time. Current approaches such as MR-SIGMA consequently rely on selecting images from pre-computed respiratory libraries, preventing visualisation of motion not seen during the pre-scan and limiting true real-time adaptation. Here, we propose a deep-learning CycleGAN-based model to reconstruct 3D image frames from real-time volumetric data with latencies short enough for adaptation.
Methods
Free-breathing abdominal scans were conducted on six healthy adult volunteers on a 3-T MRI scanner (MAGNETOM Skyra, Siemens Healthcare), using a radial stack-of-stars sequence. Each volunteer was scanned twice; first with a fully sampled stack comprising 64 spokes (185 ms per stack), and again with an accelerated sequence of undersampled stacks (22 spokes, 83.8 ms per stack). A CycleGAN model was trained to generate MR-SIGMA images given the input of a NuFFT reconstruction of three retrospectively undersampled consecutive stacks. The CycleGAN was tested on the final 50 frames of the fully sampled acquisitions (withheld from training) and applied to prospectively accelerated acquisitions.
Results
Applied to retrospectively undersampled data, the proposed CycleGAN technique produced 3D image frames with quality matching MR-SIGMA (MSE: (1.82±0.66)×10-4, SSIM: 0.965±0.005). For prospectively accelerated acquisitions, the model generated image frames from data acquired in 251 ms. Image quality was comparable to MR-SIGMA without being restricted to pre-scan bin images. Motion in the images was observed to match the motion in heavily undersampled real-time NuFFT reconstructions.
Conclusion
We have demonstrated the use of a CycleGAN model to rapidly reconstruct high-quality 3D abdominal MRI image frames with latency short enough to allow real-time tracking. Implementation of this approach in real-time on an MRI-Linac will facilitate low-latency volumetric image guidance for radiotherapy beam adaptation.