MRI is becoming a prominent imaging technique in the field of external beam radiotherapy, for both radiation therapy treatment planning and, with the recent introduction of the MR-Linac¹, for guidance during radiation. However, geometric accuracy is vital for correct delineations pre-treatment and accurate tumor tracking during radiation. Previously, we have implemented a motion model approach for dose accumulation mapping and tumor tracking during radiotherapy treatment². The method determines the current 3D motion state of the abdomen, based on 1) a statistical motion model derived from a retrospectively sorted 4D-MRI acquired prior to treatment³, and 2) fast 2D cine-MR images acquired during the treatment. The current motion state is estimated by co-registering the incoming 2D data to a 3D volume that is deformed based on the predetermined statistical model⁴. For an accurate calculation of the accumulated dose, it is essential that the estimated motion states are geometrically accurate. Here, we quantify the geometric fidelity of 1) the retrospectively sorted 4D-MRI and 2) the 2D cine-MR images using an MR-compatible motion phantom. This was achieved by moving the phantom using a range of user-defined waveforms while acquiring 4D-MRI and 2D cine-MR data and comparing the 4D and 2D positions with reference positions.

Acquisitions: All acquisitions were performed on a 1.5T MR scanner (Ingenia, Philips, Best, Netherlands). 4D-MRI data were acquired using a 3D RF-spoiled GRE with radial in-plane sampling (parameters in Table 1) interleaved with a navigator. Moreover, two orthogonal (coronal and sagittal) 2D cine-MRI slices (see Fig. 1c) were acquired in an interleaved fashion (parameters in Table 1). Additionally, static 3D volumes were acquired using the same parameters used for the 4D-MRI, which served as reference images for the various positions of the motion phantom.

Experimental set-up: An elliptically shaped motion phantom (Quasar, Modus Medical Devices Inc., London, Canada) was used that consists of two static compartments and one linearly moving cylinder, which was driven by user-defined waveforms (see Fig. 1a, b). The static parts were filled with CuSO4-doped water, while the moving cylinder was filled with doped agarose gel. A Ping Pong ball, simulating a tumor, was filled with CuSO4-doped water and inserted in the center of the cylinder. Four 4D-MRI and nine 2D data sets were acquired, while the cylinder was moving in a 1D linear fashion using the waveforms described in Table 2. The 4D-MRI waveforms were used to test the hypothesis that the 4D-MRI acquisition depicts the average position within every phase for various motion amplitudes. The nine 2D waveforms represent all expected motions within the abdomen.

Analysis: 4D-MRIs were constructed by retrospectively sorting the 4D data into ten respiratory phases through phase binning. Based on the user-defined waveforms, the average position of the ball for every phase was calculated, which are the geometrically correct (GC) positions. Next, the GC positions within the images were calculated using the static (reference) 3D acquisitions. The difference in center of mass (COM) of these positions and the positions in the 4D-MRI data sets were calculated as a measure for the geometrical error of the 4D-MRI. Finally, rigid registration was performed for all 2D MS acquisitions and the root-mean-squared error (RMSE) and normalized RMSE (NRMSE) between the estimated translations and the user-defined waveforms were calculated.

The differences between the positions determined by the navigator and the waveforms were smaller than 1 mm. Fig. 2 displays one of the calculated 4D motion trajectories in blue. Moreover, for every phase a boxplot is plotted, which was calculated using the different positions in that phase, indicating that the 4D-MRI depicts correct positions, despite large intra-phase variation. The average differences in COM over the ten phases for the four 4D-MRI acquisitions were 1.03, 1.30, 0.96 and 1.25 mm respectively, which resembles the average geometrical error for the 4D-MRI. Fig. 3 shows a coronal view of the phantom at different positions (left) and the corresponding phases of the 4D-MRI. The red lines are a visual aid for corresponding positions. The average RMSE for the 2D MS acquisitions was 1.06 mm, corresponding to an NRMSE of 4.9%. The acquisition with the frequency sweep waveform showed that the images depict the correct position for sinusoidal motion up to 0.5Hz. Using a linear MRI-compatible motion phantom the geometric errors of a retrospectively sorted 4D-MRI (1-1.3 mm) and a 2D MS acquisition (0.6-1.5 mm), which serve as input for a motion framework, were calculated and found to be smaller than the voxel/pixel size. Future studies will focus on phantom motions in more directions, to simulate more representative in-vivo motion trajectories.