Synopsis

Keywords: Image acquisition: Fast imaging, Image acquisition: Reconstruction, Image acquisition: Machine learning

MR signature matching (MRSIGMA) enables to perform real-time 4D MRI to guide radiotherapy using a combined MR-linac system. MRSIGMA shifts the acquisition and reconstruction burden to a motion learning step, where a 4D motion dictionary of 3D motion states and corresponding motion signatures is learned for each treatment fraction. Once the 4D motion dictionary is learned, fast signature-only acquisition and matching can be performed to minimize imaging latency and obtain 3D images in less than 300ms. The lecture will discuss acquisition, reconstruction and deep learning techniques to implement signature matching for radiotherapy monitoring, adaptation and dose calculation in in real-time.

Real-time MRI techniques with low total imaging latency (including acquisition and reconstruction) can be very useful to guide treatment procedures in interventional radiology and radiation oncology. For example, the MR-linac system, which combines an MRI scanner and a linear accelerator, offers an ideal platform to use MRI to guide radiotherapy [1-2].

One of the main promises of the MR-linac is to adapt radiation delivery to tumors affected by motion in real-time to increase radiation dose in the tumor while sparing healthy tissue around the tumor. [3-4]. This goal would require performing 3D MRI with a total latency lower than 300ms to account for changes due to respiratory motion. However, current MRI technology is relatively slow to image volumetric organ motion in real time, and even with the latest acquisition and reconstruction techniques, real-time MRI is limited to two-dimensional imaging, which has suboptimal interpretation of motion and through-plane motion misregistration.

A significant effort is underway to develop real-time 4D MRI techniques that are tailored to the needs of the MR-linac and minimize imaging latency. One of these techniques is MR signature matching (MRSIGMA), which consists of two steps: 4D motion dictionary learning and signature matching [5-7]. The learning step continuously acquires radial k-space data for about 5 minutes and reconstructs a 4D motion dictionary with entries given by pairs of 3D motion states and motion signatures. Motion signatures are given by the motion range along the z dimension, which is the main direction of motion in the body. After motion learning is completed, MRSIGMA moves to a real-time signature matching step, where the same acquisition is employed, but for each angle, a motion signature is computed and matched to one of the motion signatures and its corresponding 3D motion state in the pre-computed 4D motion dictionary. The signature matching step runs with fast matching operations without the need of image reconstruction, enabling real-time 4D MRI with latency lower than 300ms.

This lecture will discuss k-space acquisition, image reconstruction and the application of deep learning to implement signature matching for real-time MRI-guided radiotherapy on a MR-linac system.