Motion-Adaptive Temporal Resolution for Radial Real-Time Imaging at a Low-Field MR-Linac
Florian Friedrich1,2, Philipp Mann3,4, C. Katharina Spindeldreier5, Peter Bachert1,2, Mark E. Ladd1, Sebastian Kl├╝ter5, and Benjamin R. Knowles1

1Medical Physics in Radiology, German Cancer Research Center (DKFZ), Heidelberg, Germany, 2Department of Physics and Astronomy, Heidelberg University, Heidelberg, Germany, 3Department of Medical Physics in Radiation Oncology, German Cancer Research Center (DKFZ), Heidelberg, Germany, 4National Center for Radiation Research in Oncology (NCRO), Heidelberg Institute for Radiation Oncology (HIRO), Heidelberg, Germany, 5Department of Radiation Oncology, University Hospital of Heidelberg, Heidelberg, Germany


Hybrid MRI linear accelerators (MR-linacs) enable real-time image guidance during radiotherapy. Under real-time MRI, a compromise must be found between spatio-temporal resolution and SNR, whereas to precisely track tumour position, both should be maximised. The presented method implements a motion-adaptive image reconstruction based on a golden-angle radial acquisition scheme. This allows producing SNR-optimised images under periods of small motion and images optimised for temporal resolution when motion is larger. The technique was implemented at a low-field MR-linac (0.35T) to image a free-breathing volunteer and a motion phantom.


A hybrid MRI linear accelerator (MR-linac) enables real-time image guidance for tumour position tracking during irradiation. The aim is to either gate or update the x-ray beam position based on tumour location to reduce the radiation dose of healthy tissue and to enable tumour dose escalation. Real-time imaging is particularly relevant for treatments in the abdomen and thorax, where organ motion is particularly problematic. An additional challenge is the small signal due to a low B0 field on the MR-linac system used in this study. The aim of this work was to apply a motion-adaptive temporal resolution technique on a MR-linac, using a self-navigated golden-angle radial acquisition approach, to increase SNR for periods of small motions and temporal resolution for large motion.


All experiments were performed on a 0.35T MRIdian system (ViewRay Inc., Cleveland, Ohio, USA), using a 12-channel receive coil. A motion tracking sequence was implemented based on balanced steady-state free precession (bSSFP) [1] with a radial k-space sampling scheme using radial lines (spokes) acquired in a golden angle (Ψ1=111.24...°) [2] or tiny golden angle (Ψ10=16.95...°) [3] fashion. The sequence was used in conjunction with an offline motion-adapted reconstruction [4] implemented in Python 3.6.1. Individual images were reconstructed using the ‘sliding window’ technique in which the previous n radial spokes were used for reconstruction, using a non-uniform fast Fourier transform [5] implemented in pynufft [6]. The sliding window width (i.e., the size of n) was automatically derived from the magnitude of the detected motion as derived from observation of the magnitude of the k-space centre signal value, smoothed by a Savitzky-Golay filter before motion analysis. The minimum window width was set to 28 spokes per image, which represents a temporal resolution of 2 frames/s. Experiments were performed on a motion phantom (Fig.1) and one healthy subject. Imaging parameters of the tracking sequence were: 2D acquisition; 1 slice; TR/TE=9.0/4.5ms; (Δx)³=2.2x2.2x8mm³; FOV=280x280mm²; FA=60°; bandwidths of 200 and 800Hz/px were tested. The in-house-built motion phantom [7] was driven by a CIRS motor (CIRS Inc., Norfolk, Virginia, USA). The motor was driven with a sin4 trajectory to simulate breathing motion (3cm peak-to-peak amplitude, 6s breathing circle). Rigid image registration was performed using phase correlation [8] implemented in skimage 0.13.1, and in the case of the phantom, the mean deviation to the motion was compared between adaptive and fixed temporal resolution. The healthy subject was given no breathing instructions.


The comparison between the acquisition schemes with the standard golden and the tiny golden angle at 200Hz/px bandwidth shows a significant reduction of eddy current related artefacts for the smaller angle increment (Fig.2). One large banding artefact is visible in both images independent of the angle increment. This banding artefact disappeared by increasing the bandwidth to 800Hz/px (Fig.2), which comes at a cost of decreased SNR. The motion-adaptive temporal resolution created a large reconstruction window towards the end of the exhaled and inhaled states, in which motion is slow (Fig.3). This enabled an increase in image quality due to the increased number of spokes in the image reconstruction. During fast motion, a short reconstruction window ensured blurring was avoided (Fig.4). The image registration showed improved position detection when the adaptive reconstruction was used (Fig.5).


The tiny golden angle increment produces superior image quality compared to the standard golden angle, most likely due to the reduction of eddy current effects. The choice for an optimal bandwidth is a compromise between reducing B0-related banding artefacts and increasing SNR, especially because of the inherently low SNR of the imaging system (B0=0.35T). Moreover, a high bandwidth enables shorter gradients and a smaller repetition time and thus a higher temporal resolution, which is an advantage for reducing image motion. While the filtering of the amplitude in k-space centre and the image reconstruction were performed retrospectively, both were implemented in a way to allow real-time online imaging.


A tracking sequence with a motion-adaptive reconstruction window has been applied to a low-field MR-linac. This technique has the advantage over fixed temporal resolution approaches in that image quality is improved in areas of slow motion whilst a high temporal resolution is maintained when necessary. An advantage of radial readout over Cartesian is the continuous motion monitoring, updated by every spoke, which is much faster than motion monitoring and radiation gating based on the reconstructed Cartesian image. Future work will be focused on optimising the adaptive reconstruction window with regards to image quality for tracking algorithms as well as more advanced real-time data reconstructions to improve image quality.



1. Scheffler K, et al. Principles and applications of balanced SSFP techniques. European Radiology 2003;13(11):2409-2418.
2. Winkelmann S, et al. An optimal radial profile order based on the golden ratio for time-resolved MRI. IEEE Transactions on Medical Imaging 2007;26(1):68-76.
3. Wundrak S, et al. Golden ratio sparse MRI using tiny golden angles. Magnetic Resonance in Medicine. 2016;75(6):2372-2378.
4. Friedrich F, et al. Self-Adapting Dynamic Temporal Resolution for Optimized Reconstruction of Free-Breathing Radial Real-Time MRI. ESMRMB 2017, Magnetic Resonance Materials in Physics, Biology and Medicine 2017;30(S1):292.
5. Fessler J, et al. Nonuniform fast Fourier transforms using min-max interpolation. IEEE Transactions on Signal Processing 2003;51(2):560-574.
6. Lin J, et al. Pynufft: python non-uniform fast Fourier transform for MRI, Building Bridges in Medical Sciences 2017, St John's College, CB2 1TP Cambridge, UK, 2017.
7. Leiner L, et al. Planning and implementation of a phantom for MR-guided therapy [in German only]. Bachelor Thesis, Department of Physics and Astronomy, Heidelberg University, Germany, 2017.
8. Guizar-Sicairos M, et al. Efficient subpixel image registration algorithms. Optics Letters 2008;33:156-158.


Figure 1: Setup of the motion phantom and the motor on the couch of the MR-linac. The cylinder (72 mm diameter, 200 mm length) was moved inside the phantom to simulate breathing motion. Inside the cylinder was a mixture of agarose and gadolinium contrast agent to mimic liver tissue as well as two additional contrast gels in hollow plastic spheres.

Figure 2: Impact of different golden angle increments and bandwidths on the artefacts in the phantom images. Top: The golden angle Ψ1 led to artefacts marked with the arrows that were produced by eddy currents due to rapidly changing gradients. The use of tiny golden angle increment Ψ10 attenuated gradient changes nearly eliminated the artefacts. Bottom: A higher bandwidth decreased the SNR but also decreased artefacts due to B0 inhomogeneity (banding artefact marked with arrow). All images had a reconstruction window of 300 spokes.

Figure 3: The amplitude in k-space centre was smoothed by a Savitzky-Golay filter and used for respiration monitoring. Maxima and minima represent the exhaled and inhaled state of the breathing cycle when only very little motion is present. For every new image frame the motion-adaptive temporal resolution combines all adjacent previous spokes whose deviation in signal intensity is within a predefined threshold. This leads to a large reconstruction window width during little motion to increase the image quality and to a small reconstruction width during fast motion when a high temporal resolution is needed.

Figure 4: Sagittal images of the abdomen of a freely breathing healthy volunteer (top) and the motion phantom (bottom). The reconstruction window width ranged from 28 up to 146 for the in-vivo images and from 28 to 167 for the motion phantom images. The wider reconstruction window led to increased image quality due to increased SNR and decreased under-sampling artefacts (left). The narrower reconstruction window (middle) was needed during motion to avoid image blurring (right). All images were acquired with a 200 Hz/px bandwidth and a tiny golden angle increment (Ψ10 = 16.95...°).

Figure 5: Rigid image registration of the contrast circle in the motion phantom. During the lower plateaus with little motion, the registration was more accurate when the adaptive reconstruction technique with 28-167 spokes was used (blue). The fixed reconstruction window of 28 spokes (orange) always generated under-sampling artefacts, which complicated image registration. A fixed number of 167 spokes (red) results in a delay towards the motor motion due to image blurring. The mean deviation of the adaptive temporal resolution was 1.2 pixel and thus smaller than for the fixed resolutions with 1.8 pixel (28 spokes) and 2.4 pixel (167 spokes).

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)