Room: Track 2
Purpose: Increased soft-tissue contrast on MRI-Linacs improves tumour delineation during radiation therapy. Techniques that can dynamically adapt the radiation beam to tumor motion will lead to more accurate MRI-Linac treatments that reduce collateral healthy tissue damage. However, compressed sensing reconstruction of undersampled MR data is computationally slow, presenting a barrier to real-time adaptation. Here, we use automated transform by manifold approximation (AUTOMAP), a generalized framework that maps raw MR signal to the target image domain, to rapidly reconstruct images from an MRI-Linac with machine learning.
Methods: AUTOMAP was implemented in TensorFlow with an architecture consisting of 2 dense, 2 convolution and 1 deconvolutional layers. Datasets of 80,000 training images and 4,000 validation images were sourced from ImageNet and encoded with MATLAB to raw signal data for cartesian/radial MRI acquisitions. Raw k-space data of the thorax of a healthy volunteer were acquired with a cartesian gradient echo sequence on an MRI-Linac and exported for offline reconstruction. K-space data for a Golden-angle radial acquisition, a benchmark for motion-sensitive imaging, was simulated using the digital CT/MRI breathing XCAT (CoMBAT) phantom and used to compare AUTOMAP to conventional compressed sensing reconstruction techniques.
Results: Training of the machine-learning-based reconstruction framework took 16 hours on a 12 GB GPU. AUTOMAP successfully reconstructed cartesian k-space data into 128×128 cine images (20 frames) of the human thorax in a single forward pass (10 ms/image reconstruction time). Digital phantom tests showed that the reconstruction accuracy of AUTOMAP (Root-mean-square error, RMSE 3.3%) for 4x undersampled radial data was close to wavelet techniques (RMSE 3.2%) whilst being nearly 50 times faster to execute (47 ms vs 2210 ms).
Conclusion: We have demonstrated rapid reconstruction of k-space data from an MRI-Linac with AUTOMAP. Planned real-time implementation of these techniques promises to reduce reconstruction latency for MRI-Linac treatments with dynamic tumor tracking.
Funding Support, Disclosures, and Conflict of Interest: Funding Support: This work has been funded by the Australian National Health and Medical Research Council Program Grant APP1132471. D Waddington is supported by a Cancer Institute of NSW Early Career Fellowship 2019/ECF1015.