Accurate localization of moving targets, such as liver and lung cancer, is critical for ensuring precise delivery of radiation therapy. This is especially crucial for stereotactic body radiation therapy (SBRT) where high radiation dose is delivered in each fraction. The only way to achieve absolute precision in localizing moving targets is through real-time 3D imaging of the target during the treatment delivery.

MRI integrated with a radiotherapy unit has been under intensive development recently for both inter- and intra-fraction verification. Compared to CBCT, MRI has much better soft tissue contrast and no radiation dose. Currently, commercial MRI-Radiotherapy systems use 2D-cine MR images for real-time imaging of lung tumors for gated treatment.¹ However, it cannot capture the out of plane motion of the target due to the lack of volumetric information. 4D-MRI is under development through either prospective or retrospective approaches. Due to limitations of hardware and software, prospective 4D-MRI suffers from poor temporal resolution (~1 s) and poor spatial resolution (4-5 mm).^2-3 Retrospective 4D-MRI has better temporal and spatial resolution, but suffers from long acquisition time (5-30 min).^3-6 No real-time volumetric cine MRI has ever been developed due to the limitation of the MR data acquisition speed.

The purpose of this study is to develop an ultrafast volumetric cine MRI (VC-MRI) technique for real-time 3D target localization and tracking of liver and lung radiotherapy.

Patient 4D MRI data acquired at the radiotherapy planning stage are used as the patient prior images. Principal component analysis (PCA) is used to extract 3 major respiratory deformation patterns,$$$\{ \tilde{D}^j_0\}$$$, from the deformation field maps (DFMs) registered between end-expiration phase and all other phases of the 4D MRI. The on-board VC-MRI at any instant is considered as a deformation of the prior MRI at the end-expiration phase. The deformation field for VC-MRI is represented by a linear combination of the 3 major deformation patterns.⁷ The weightings of the deformation patterns are solved by minimizing the data fidelity function in Eq. 1, which minimizes the difference between the corresponding 2D slice of the VC-MRI and the on-board 2D-cine MRI acquired.

$$$f(w) = \left\Vert S*VCMRI\left(D_{0ave}+\sum_{j=1}^3 w_j\tilde{D}^j_0, MRI_{prior}\right)-2DCine_{slice} \right\Vert^2_2$$$ Eq. 1

Here, S is the operator to extract the corresponding 2D slice from the estimated VC-MRI images, D_0ave is the average of the original DFMs, w_j are the weightings corresponding to each principal motion mode.

The ultrafast on-board 2D-cine images were acquired by sampling only about 10% of the k-space. 85% of the undersampled k-space data were taken at the central k-space and the other 15% were randomly sampled elsewhere. An iterative MR reconstruction algorithm with a total generalized variation (TGV) penalty was used to reconstruct the cine image from the undersampled 2D k-space data. This extremely low sampling rate potentially allows ultrafast 2D-cine acquisition of 20-30 frames/s. Correspondingly, VC-MRI can be generated at 20-30 frames/s when a single 2D-cine image is used for each VC-MRI estimation.

The VC-MRI method was evaluated using an XCAT (computerized patient model) simulation of a lung cancer patient and 4D-MRI data from a liver cancer patient. In the XCAT study, a breathing amplitude decrease of 35% was simulated from prior 4D-MRI to on-board cine MRI images. In the liver cancer patient data, a breathing amplitude increase existed from prior to on-board images. The accuracy of the estimated VC-MRI was quantitatively evaluated using Volume Percent Difference(VPD), Center of Mass Shift(COMS) and normalized cross correlation(NCC). Effect of cine acquisition orientation on the estimation accuracy was also evaluated.

VC-MRI can be generated with a frame rate as high as 20-30 frames/s. Figure 1 shows the undersampled 2D-cine images simulated by sampling 10% of the k-space for both XCAT and patient data. Figures 2 and 4 show the VC-MRI estimated using the ultrafast undersampled 2D-cine images for XCAT and patient data, respectively. Figures 3 and 5 show the corresponding subtraction images. Excellent agreement was achieved between estimated and ground-truth VC-MRI. The NCC between the estimated and ground-truth ultrafast VC-MRI in the axial, coronal and sagittal views were on average 0.952, 0.955, 0.926 for all XCAT data and 0.980, 0.937, 0.988 for patient data. Among the XCAT data, the maximum VPD between ground-truth and estimated lesion volumes in the ultrafast VC-MRI was 6.77% and the maximum COMS was 0.96 mm. Estimation using axial/sagittal orthogonal cines yielded slightly better accuracy than using axial cines only. Preliminary studies showed for the first time that it is feasible to generate ultrafast high quality VC-MRI up to 30 frames/s for real-time 3D target localization before or during radiotherapy treatments.