INTRODUCTION

Quantitative dynamic contrast-enhanced (DCE) MRI is a powerful diagnostic and monitoring tool for treatment response of cancers such as pancreatic ductal adenocarcinoma (PDAC). MR Multitasking has shown that a low-rank tensor image model can achieve high spatial and temporal resolution for imaging PDAC^1,2 at the standard Gd dose. Low-dose imaging is attractive for longitudinal monitoring, but reduced signal-to-noise ratio warrants more efficient image models than low-rank tensors alone. Non-rigid motion compensation (MoCo) can complement low-rank image models to more efficiently represent respiratory motion^3-5, but computation is challenging for large-scale problems. To address these limitations, we integrated non-rigid MoCo into the MR Multitasking framework, directly applying motion fields to spatial factors (eigenimages) to reduce computational cost by 3,000×.

METHODS

Pulse sequence: A 3D FLASH sequence with periodic nonselective saturation-recovery (SR) preparation used an axial golden-angle stack-of-stars sampling scheme with Gaussian k_z density to collect imaging data⁶. A center k-space line in the superior-inferior direction was acquired every 8 readouts as training data.
Image reconstruction: Figure 1 illustrates the reconstruction pipelines for standard Multitasking (No-MoCo) and MoCo Multitasking. In both pipelines, a “real-time” (ungated) basis ($$$\mathbf{\Phi}_{\textrm{rt}}$$$) was generated from the SVD of the training data ($$$\mathbf{d}_{\textrm{rt}}$$$) and used for preliminary estimations of spatial factor $$$\mathbf{U}_{\textrm{rt}}$$$ from the imaging data ($$$\mathbf{d}_{\textrm{img}}$$$). The reconstructed image $$$\mathbf{A}_{\textrm{rt}}=\mathbf{U}_{\textrm{rt}}\mathbf{\Phi}_{\textrm{rt}}$$$ was binned into six respiratory states (bin$$${}_{n}$$$, $$${n}$$$=1,2,…,6).
For MoCo Multitasking, each motion field $$$\mathbf{M}_{n}$$$ mapping the end-expiratory bin (bin₁) to bin$$${}_{n}$$$ was estimated from the respiratory bin average images using non-rigid registration in ANTs⁷. Motion-compensated training data ($$$\mathbf{d}_{\textrm{MoCo,tr}}$$$) were generated by applying the inverse motion fields to $$$\mathbf{U}_{\textrm{rt}}$$$ and re-encoding the central k-space line. A multidimensional tensor subspace basis was then estimated using low-rank tensor completion as in our prior work¹, either from $$$\mathbf{d}_{\textrm{tr}}$$$ (for No-MoCo Multitasking) or from $$$\mathbf{d}_{\textrm{MoCo,tr}}$$$ (for MoCo Multitasking). The spatial factor $$$\textbf{U}$$$ was recovered according to
$$ \textbf{U}=\underset{\textbf{U}}{\textrm{argmin}}\sum_{n}^{ }\left\| \mathbf{d}_{\textrm{img},n}-\Omega _{n}\left ( \mathit{\Phi_{n}}\times _{1} \left [ \textbf{FSM}_{n}\mathbf{U} \right ] \right )\right\|_{2}^{2}+R(\textbf{U})$$
where $$$\Omega _{n}$$$ applies (k,t)-space undersampling corresponding to $$$\mathbf{d}_{\textrm{img},n}$$$, the subset of imaging data at motion state $$${n}$$$,$$$\mathit{\Phi}_{n}$$$ is a subtensor the temporal factor at bin $$${n}$$$, $$$\textbf{F}$$$ takes the Fourier transform, $$$\textbf{S}$$$ applies coil sensitivities, and R(·) is wavelet sparse regularization. No-MoCo Multitasking solved the same problem with $$$\mathbf{M}_{n}=\mathbf{I}$$$. The final multidimensional image tensor $$$\textit{A}$$$ was computed as $$$\textit{A} = \mathit{\Phi}_{n}\times_{1}\mathbf{U}$$$.
Imaging experiments: All data were acquired on a 3T scanner (VIDA, Siemens) with the following parameters: TE/TR = 2.2/6.1ms, SR period = DCE temporal resolution=1030ms, Flip angle=8°, scan time=10min, acquisition matrix=224×224×50 with 15% oversampling in k_z direction, spatial resolution=1.3×1.3×4.0–1.5×1.5×4.0mm³. A 20%-dose (0.02mmol/kg) of Gadavist contrast agent was administered 2 minutes into the scan at the rate of 2ml/s. Five healthy volunteers were enrolled. Liver-dome positions were tracked across the DCE dimension within the end-expiratory bin (bin₁) to measure intra-bin motion. A one-tailed paired t-test tested the hypothesis that MoCo reduced the standard deviation of liver-dome position within bin₁.

RESULTS

Figure 2 shows the images of the arterial phase (~200s) from No-MoCo and MoCo Multitasking. The MoCo Multitasking high-order singular values along the respiratory dimension decay more quickly than No-MoCo Multitasking, indicating a lower effective rank. MoCo reduced variation between bin₁ (end-expiration) and bin₆ (end-inspiration) as intended.
Figure 3 shows the contrast and position variations of a 1D liver-dome profile along the DCE dimension in bin₁. The 1D profile from MoCo Multitasking has a more stable liver-dome position than that from No-MoCo Multitasking.
Fig. 4 shows the detected liver-dome position within bin₁ for all reconstructions, along with standard deviations. MoCo Multitasking reduced liver-dome position variation (p=0.009), confirming reduced intra-bin motion.
Figure 5 shows quantitative pancreatic DCE maps for a representative subject. Unreasonable values and abrupt changes appear on the DCE maps from No-MoCo Multitasking. The MoCo images are more homogeneous within the pancreas (as expected for healthy subjects) and produced average V_e, K^trans, and V_p values more consistent with those reported in our previous standard-dose DCE studies^1,2.

DISCUSSION

Directly applying motion fields to eigenimages rather than time-domain images makes non-rigid MoCo practical for high-dimensional MR Multitasking, with 3,045× fewer Fourier transforms and motion deformations. Non-rigid MoCo reduced the rank in the respiratory dimension, indicating more efficient modeling for respiratory motion, crucial for low-dose DCE imaging. MoCo reduced intra-bin motion in the end-expiratory bin and showed potential for improved DCE quantification. The potential for reliable low-dose free-breathing DCE application warrants further study in a larger cohort, including patients, assessing repeatability/reproducibility, and comparing with standard-dose DCE MRI.

CONCLUSION

Non-rigid MoCo can be applied directly to eigenimages in MR Multitasking, reducing intra-bin motion in low-dose free-breathing abdominal DCE imaging.

Acknowledgements

This work was partially supported by NIH R01 EB032801.