Imaging latency of Golden Angle 2D MRI using real-time GPU-accelerated image reconstruction for MR-guided radiotherapy
Pim Borman1, Bas Raaymakers1, and Markus Glitzner1

1Radiotherapy, UMC Utrecht, Utrecht, Netherlands


For tumor tracking in MRI-guided radiotherapy it is important to minimize the latency between the moment of anatomic change and its appearance on the MR image. By using a 2D golden angle sampling trajectory in combination with a sliding window reconstruction, the latency can be decoupled from the frame rate, yielding frame rates of up to one repetition time. We implemented a real-time GPU-accelerated reconstruction pipeline where k-space data is directly streamed to a reconstruction server during acquisition. Using this we investigated the influence of the sliding window width on the latency, reconstruction time, frame rate and image quality.


Hybrid MR-Linac systems will facilitate real-time treatment adaption by continuous beam steering based on tumor tracking to compensate for e.g. respiratory motion. For this it is important to maximize the MR imaging frame rate and to minimize the latency between the moment of anatomic change and the reaction of the multi-leaf collimator (MLC). MR imaging is the major contributor to the total latency of this feedback chain.1,2 In this work we implemented a real-time GPU-accelerated reconstruction pipeline for a 2D golden angle (GA) tracking sequence with the aim of minimizing reconstruction latency. The GA sampling scheme enables the use of a sliding window reconstruction to increase the frame rate and essentially decouple it from the latency. The window width (i.e. number of profiles per frame) can be adjusted dynamically, and directly influences the imaging latency and image quality.


The experiments were performed on a 1.5T MR-Linac system (Elekta, SE), using a 2D RF-spoiled GRE sequence with a GA radial profile ordering (TR/TE: 3.7/1.8 ms). 1D sinusoidal motion was generated using a 4D motion phantom (ModusQA, CA) set to 3 s period and 15 mm amplitude. The absolute positions were recorded on an external workstation and served as the reference. A custom library was developed and added to the system’s reconstruction framework Recon 2.0 (Philips, NL) to continuously stream k-space profiles to the external workstation, during acquisition. From there, the profiles were forwarded to a reconstruction server equipped with a TITAN Xp GPU (NVidia, CA, US), where a non-uniform FT was performed.3 The resulting images were send back to the external workstation where they were provided with a timestamp and saved. The number of profiles per reconstructed image could be changed on-line while scanning by pressing ‘u’ (up) for more or ‘d’ (down) for less profiles. Using this pipeline, one minute acquisitions were made for window widths of 50, 75, 100, 150, 250, and 300 profiles. For each, the latency was calculated by fitting a sinusoidal model to the moving center-of-mass of the high-contrast object and to the reference. The image quality was quantified by calculating the artifact power caused by streaking.4 Additionally, reconstruction times, and frame rates were calculated.

Results and Discussion

The results show a linear increase of latency (Fig. 2a) with number of profiles per frame, ranging from 143 ms to 665 ms for 50 and 300 profiles, respectively. This includes acquisition, reconstruction, and data handling latencies. Earlier work has shown that for radial trajectories the acquisition latency is equal to ½ the acquisition time [1], while the reconstruction latency is shown in figure 2b. With this, the total imaging latency consists of 65% - 84% acquisition, 6% - 15% reconstruction, and 10% - 20% data handling, depending on the number of profiles. Although the frame rate (Fig. 2c) could in principle be 1/TR, independent of the number of profiles, in our implementation it was restricted by the reconstruction and data handling times, which limited the maximum sliding window factor. As expected, the artifact power decreases exponentially with the window width (Fig. 2d), decreasing the amount of streaking significantly (Fig 3). Depending on the application, there is a trade-off between latency and image quality. With our implementation this trade-off could be made on-line by dynamically changing the window width.


Our experiments showed that by using a GPU-accelerated real-time reconstruction pipeline, the contributions of reconstruction and data handling to the total imaging latency are minor compared to the contribution of the MR acquisition itself, making GA sliding window acquisitions suitable candidates for tracking. Furthermore, our implementation allowed for a dynamic trade-off between latency and image quality by tuning the window width during acquisition.


This work was supported by the ITEA Starlit grant (project 16016).


[1] Borman et al., Characterization of imaging latency for real-time MRI-guided radiotherapy. PMB 2018; 63(15).

[2] Glitzner et al., On the suitability of Elekta's Agility 160 MLC for tracked radiation delivery: closed-loop machine performance. PMB 215; 60(5).

[3] Knoll et al., An Open-Source GPU Library for 3D Gridding with Direct Matlab Interface. Proc. ISMRM 2014

[4] Xiao et al., Comparison of Parallel MRI Reconstruction Methods for Accelerated 3D Fast Spin-Echo Imaging. MRM 2008; 60(3).


Figure 1: Illustration of the setup. The GA acquisition is started from the MR Host and during acquisition the k-space data is streamed to an external workstation. Simultaneously the positions of the 4D motion phantom are streamed to the same workstation. The k-space data is forwarded to a reconstruction server equipped with a GPU, and images are send back and stored. The sliding window width is adjusted by pressing ‘d’ or ‘u’. Afterwards, the latency is calculated.

Figure 2: Measured latency (a), frame rate (b), reconstruction time (c), and artifact power (d) as a function of the sliding window width, i.e. number of profiles per reconstructed image frame. Acquisition time per profile equals TR (3.7 ms).

Figure 3: Selected frames of the moving phantom for several sliding window widths of 50 until 300 profiles. Increasing the window width reduces the streaking artifacts but increases the latency.

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