Real-Time MR-Based 3D Motion Monitoring Using Raw K-Space Data

Abstract

Due to its great soft-tissue contrast and non-invasive nature, magnetic resonance imaging (MRI) is uniquely qualified for motion monitoring during radiotherapy. However, real-time capabilities are limited by its long acquisition times, particularly in 3D, and require highly undersampling k-space resulting in lower image resolution and image artifacts. In this paper, we propose a simple recurrent neural network (RNN) architecture to continually estimate target motion from single k-space spokes. By directly using the incoming k-space data, additional image reconstruction steps are avoided and less data is required between estimations achieving a latency of only a few milliseconds. The 4D XCAT phantom was used to generate realistic data of the abdomen affected by respiratory and cardiac motion and a simulated lesion inserted into the liver acted as the target. We show that using a Kooshball trajectory to sample 3D k-space gives superior results compared to a stack-of-stars (SoS) trajectory. The RNN quickly learns the motion pattern and can give new motion estimations at a frequency of more than 230 Hz, demonstrating the feasibility of drastically improving latency of MR-based motion monitoring systems.