Introduction

Accurate radiation beam delivery is vital in radiation therapy (RT), especially for treatment techniques with large dose fractions, such as stereotactic body radiation therapy (SBRT). Real-time tumor motion tracking is an emerging technique for monitoring radiation delivery, and its benefits on radiation dosimetry have been proved¹. In magnetic resonance-guided Radiotherapy (MRgRT), current onboard imaging relies on single or orthogonal two-dimensional (2D) cine magnetic resonance imaging (MRI)², which only provides T1-weighted (T1w) image contrast and incomplete tumor motion. Multi-parametric (Mp) four-dimensional (4D) MRI provides versatile tumor contrasts and accurate tumor motion; however, the imaging time is prohibitively long for real-time application³. According to the American Association of Physicists in Medicine (AAPM) task group report, real-time tumor tracking requires the total latency between a respiratory motion and dose delivery to be less than 500 ms⁴, allowing an even shorter time for image acquisition and reconstruction. This study developed a deep learning-based dual-supervised downsampling-invariant deformable registration (D3R) model for motion estimation from ultra-undersampled radial k-space and reconstructed high-quality (HQ) Mp 4D-MRI in real-time. Our results showed that the D3R model is robust to incomplete k-space data up to a downsampling factor of 500. The real-time HQ Mp 4D-MRI accurately reflects the tumor motion trajectory with sub-voxel level accuracy.

Methods

Thirty-nine patients were collected from Beijing Cancer Hospital with institutional review board approval. Each patient underwent a T1w 4D-MRI using TWIST volumetric interpolated breath-hold examination (TWIST-VIBE) sequence⁵. Of the thirty-nine patients, twenty-seven also underwent regular T1w and T2-weighted (T2w) 3D MRI scans. Details of the imaging parameters are listed in Table 1. For D3R model training, the original 4D-MRI were sorted into 10 phase bins based on the body area⁶. After data cleaning, the training and validation sets comprised 120 image pairs from 24 patients, and the testing set comprised 15 image pairs from 3 patients. Image binning was not needed for later HQ Mp 4D-MRI reconstruction, as deformations were directly estimated on the time-resolved original 4D-MRI.
Figure 1 illustrates the data preparation and D3R model training processes. As shown in Figure 1a, the original 4D-MRI images were retrospectively down-sampled using non-uniform fast Fourier transform (NUFFT) in 3D radial read-out. Typical linear accelerator response time is 80 ms⁷, and GPU-based 3D NUFFT reconstruction takes 50 ms, leaving around 300 ms for image acquisition. Since typical time of repetition (TR) in the T1w MRI sequence is around 3 ms, 100 radial spokes were utilized to reconstruct the training images. To evaluate the robustness of the D3R model, test images with 25, 50, 100, 150, 200, and 400 radial spokes were also included. The downsampling factor was defined as R=(M_x×M_y×π/2)/N_SP, where N_SP is the number of radial spokes and M_x and M_y are matrix size in x and y directions, respectively. Figure 1b shows the D3R model training process, which was based on 3D U-Net⁸. Reference deformation vector fields (DVFs) were calculated between the original image pairs using parametric total variation (pTV) algorithm. The training was dually supervised by the end-to-end point error (EPE) between reference DVFs and predicted DVFs and the normalized cross-correlation (NCC) between warped images and fixed images. In testing, both EPE and image similarity metrics, including structure similarity index (SSIM), peak signal-to-noise ratio (PSNR), and mean squared error (MSE), were calculated between the warped and fixed images. Conventional methods such as Elastix, Demons, and pTV algorithms were included in the performance evaluation for comparison.
Figure 2 illustrates the HQ Mp 4D-MRI reconstruction process. 4D-DVF obtained from downsampled 4D-MRI via D3R model was applied to the pre-aligned 3D MRI to reconstruct HQ Mp 4D-MRI at corresponding frames. For motion accuracy validation, tumor trajectories were tracked and compared both in the original 4D-MRI and reconstructed HQ Mp 4D-MRI.

Results

Figure 3 plots the evaluation metrics versus downsampling factor R. As shown in Figure 3a, the Demons and pTV algorithms crashed and gave increasingly different DVFs compared to those calculated on the original images as R becomes larger. Figure 3b, 3c, and 3d demonstrate similar results, in which the image similarity drops fast as R grows. By contrast, the proposed D3R model gives consistent DVFs and image similarity at R as high as 500. Figure 4 shows an example of the reconstructed HQ Mp 4D-MRI. Although the image quality was greatly degraded due to downsampling (Figure 4b), the reconstructed high-quality 4D-MR images still show detailed structures (Figure 4c, 4d) and accurate motion trajectory (Figure 4e). The average lesion motion errors are 1.37 mm and 1.23 mm for T1w and T2w HQ 4D-MRI among all the frames, respectively. D3R model inference takes about 60 ms, making the whole process within 500 ms.

DISCUSSION & CONCLUSION

We successfully obtained HQ Mp 4D-MRI in real-time via the motion information estimated from ultra-downsampled radial k-space using the D3R model. The predicted DVFs from the D3R model are consistent for the down-sampling factor up to 500, and the lesion motion trajectory was accurately reflected in the 4D-MRI. Currently, 4D-MRI is rarely used in tumor tracking for its long acquisition time and poor quality. This technique shortened the total imaging time to 500 ms and enhanced its quality, making it more practical in RT.

Acknowledgements

No acknowledgement found.