Liver dynamic contrast enhanced MRI (DCE-MRI) requires high spatiotemporal resolution to clearly visualize enhancement patterns. Image reconstruction quality is often compromised due to unavoidable respiratory motion during the long-time acquisition. This abstract presents a novel sliding motion corrected low-rank plus sparse (SMC-LS) reconstruction algorithm for free-breathing liver DCE-MRI. The sliding motion of the liver along the superior-inferior direction is estimated directly from the under-sampled k-space data. Low-rank and sparse regularization are enforced on the sliding motion corrected image reconstructions. Results demonstrated that SMC-LS can substantially reduce motion blurring and preserve more details in free breathing liver DCE-MRI.
In SMC-LS, the liver 4D DCE-MRI reconstruction
problem is formulated as
$$ \min _{L_i,S_i} \ \sum_{i=1}^{n_z}(\lambda_{L}||L_i||_* + \lambda_{S}||TS_i||_1)+\frac{1}{2}|| E( \tau \sum_{i=1}^{n_z}(R_i(L_i+S_i)))-D||_2^2 $$
where the first part exploits low rank and sparsity of the reconstructed image, and the second part enforces data consistency. $$$n_z$$$ is the number of axial slices,$$$L_i$$$ is the low-rank component and $$$S_i$$$ is the sparse component of axial slice i in the 4D image after inter-frame sliding motion registration,$$$T$$$ is the total variation operator which enforces sparsity, $$$\lambda_L$$$ and $$$\lambda_S$$$ are regularization parameters.$$$R_i$$$ is an operator that puts the motion corrected slice image,$$$L_i+S_i$$$ ,back into a 4D image. $$$E$$$ denotes the k-space data acquisition operator, which performs a multiplication by multi-channel coil sensitivities followed by an under-sampled Fourier transform, and $$$D$$$ denotes the acquired raw data.$$$\tau$$$ is a motion operator which models the frame-by-frame image transformation driven by respiration. In free breathing, the major component of liver motion is its sliding along the superior-inferior (SI) direction, whereas the abdominal wall moves a little bit inwards and outwards, and the back bone stays stationary. In SMC-LS, $$$\tau$$$ is simplified to one dimensional sliding motion of inner organs. Based on the acquired k-space center data of axial slices, inter-frame sliding motion in the SI direction is estimated by one dimensional data registration and coil clustering6. A combination of singular value thresholding used for matrix completion7 and iterative soft thresholding used for sparse reconstruction8 is used to solve the above minimization problem.
For the validation of SMC-LS, we reused one case of in vivo liver DCE-MRI data acquired in PROUD3 with single breath-hold protocol. New in vivo data were acquired in another volunteer with IRB approval using the same parameters: 1.5T scanner, 8-channel cardiac coil, multiphase spiral LAVA with 48 golden angle variable density spiral leaves per fully sampled volume, voxel size 1.48×1.48×5 mm3, and matrix size 256×256×36. A total of 4 volumes (192 spiral leaves) were continuously acquired using free breathing protocol but without contrast injection. Free breathing liver DCE-MRI data were artificially made by firstly transforming the breath-hold liver DCE MR images with the sliding motion signal derived from the free breathing case and then acquiring the k-space data with the E operator. XD-GRASP5 and conventional L+S4 reconstruction were also tested for comparison.
For L+S and SMC-LS, one frame was reconstructed using every two continuous spiral leaves, resulted in totally 96 frames at about 2 frames/s and. For XD-GRASP, two configurations were tested: 6 DCE states by 6 motion states, and 4 DCE states by 4 motion states. Fig. 1 shows a comparison of reconstructed images of L+S and SMC-LS in axial sections with zoomed-in ROIs, and Fig. 2 shows an image comparison in coronal sections with zoomed-in ROIs. Compared to the convention L+S, motion blurring artefacts were significantly reduced, and image details were better preserved by the proposed method. Fig. 3 shows a comparison of reconstructed images of XD-GRASP and SMC-LS. Compared to the proposed method, XD-GRASP reconstruction with 36 states has apparent under-sampling artefacts, while XD-GRASP reconstruction with 16 states were temporally smoothed due to lower temporal resolution.