The available acquisition time is confined to the 200-250ms diastolic period per beat to avoid cardiac motion², and single breath-hold imaging or short free-breathing imaging with navigators is needed to avoid image blurring due to respiratory motion while minimizing contrast agent washout and gross patient motion. To achieve a high resolution under these time constraints, this work presents a novel reconstruction method using COmpressed Sensing with Edge Preservation (COSEP), where the compressed sensing (CS)³ framework allows stable reconstruction from a highly undersampled dataset, while weighted total variation is locally applied to preserve edges which might be compromised using the conventional CS reconstructions. The combination of CS and edge preservation could effectively reconstruct fine details in infarcted regions, facilitating accurate characterization of scar.

In MCLE¹, the multi-contrast image series is low-rank: the temporal signal-relaxation characteristic vectors could be extracted from a training set using principal component analysis (PCA). The training set is generated using a pre-determined parametric signal model $$$M_{ss}(1-2e^{-t/T_{1}^{\ast}})$$$, where $$$T_{1}^{\ast}$$$ is a function of $$$T_{1}$$$,$$$T_{2}$$$ and flip angle⁴. As long as the parameters such as the $$$T_{1}^{\ast}$$$ range and flip angle used in the signal model match those for the acquired MCLE images, the signal recovery of each pixel could be represented with a negligible error by summing up the first few PCs with corresponding coefficients as weights. In the COSEP reconstruction, the first few PCs are utilized to recursively transform spatiotemporal signal recovery vectors to spatial principal component (PC) maps, facilitating reduction of noise and incoherent artifacts; weighted total variation (TV) regularization⁵ is locally applied to the CS framework to preserve anatomical edges in the selected regions on the spatial PC coefficient maps. The proposed COSEP reconstruction consists of finding the PC coefficient matrix $$$\mathbf{C}$$$ as the solution of the following minimization problem:

$$\arg\min_{\substack{\mathbf{C}}}\{\|\mathbf{F_{u}\Phi^{-1} C}-\mathbf{D}\|_{2}^{2}+\sum_{n}\lambda_{n}J(\mathbf{C}^{(n)})\},$$

$$\text{where}\;J(\mathbf{C}^{(n)}) = \sum\limits_{i,j}\alpha_{i,j}^{(n)}(|\mathbf{C}^{(n)}_{i+1,j}-\mathbf{C}^{(n)}_{i,j}|+|\mathbf{C}^{(n)}_{i,j+1}-\mathbf{C}^{(n)}_{i,j}|), \alpha_{i,j}^{(n)}\overset{\text{def}}{=}\frac{1}{\sqrt{1+|(\triangledown K\ast \mathbf{v}^{(n)})_{i,j}|^{2}_{2}/\beta^{2}}}$$

where $$$\mathbf{C}^{(n)}$$$ is the $$$n^{\text{th}}$$$ PC coefficient map formed using the $$$n^{\text{th}}$$$ column of $$$\mathbf{C}$$$; the operator $$$\mathbf{\Phi}^{-1}$$$ transforms the PC coefficient maps to the MCLE image series; $$$\mathbf{F_{u}}$$$ is a partial Fourier transform operator; $$$\mathbf{D}$$$ is the k-space matrix, where each column represents an undersampled k-space at a specific inversion time; $$$\{\lambda_{n}\}$$$ is a sequence of regularization parameters, which determines the trade-off between the data fidelity and the regularization; $$$\mathbf{v}^{(n)}$$$ is the initial estimate of the $$$n^{\text{th}}$$$ PC coefficient map or its further updated version; $$$K$$$ is the operator to perform weighted TV flow denoising⁵ on $$$\mathbf{v}^{(n)}$$$, which allows effective denoising without compromising edges; the spatial weight $$$\mathbf{\alpha}^{(n)}$$$ in the weighted TV regularizer $$$J$$$ is calculated on the denoised $$$\mathbf{v}^{(n)}$$$. $$$\mathbf{\alpha}^{(n)}$$$ is small at anatomical edges which are characterized by large local gradients; $$$\beta$$$ is chosen to yield small $$$\alpha$$$ at anatomical edges (e.g. $$$\beta = 0.08$$$). The nonlinear conjugate gradient descent algorithm is used to solve this problem. Once the coefficient matrix $$$\mathbf{C}$$$ is reconstructed, the MCLE image series can be obtained by a coefficient-weighted sum of PCs.

Three Yorkshire pigs with six-week-old infarcts were imaged after injection of 0.2mmol/kg Gadolinium-DTPA using an ECG-gated 3D MCLE sequence with a 160x160x10 acquisition matrix over a 1.5cm-thick slab with corresponding resolution of 1.5 mm³. The undersampled datasets were prospectively acquired using Variable Density Poisson-disk Sampling patterns at a net acceleration of 5 with a 16-channel anterior cardiac coil array in a GE 3T scanner. The data processing pipeline is shown in FIG. 1. For comparison, an alternative CS reconstruction method REPCOM⁶ was implemented and the IR-FGRE images (FIG. 3) were also acquired with the parameters as follows: matrix = 160x160; FOV = 24cm; slice thickness = 5mm. As seen in FIG. 2, COSEP provides the highest reconstructed spatial resolution on the magnitude image, computed exclusively within the area indicated by the box. Specifically, the anatomical edges in the infarct region indicated by the arrow are well-preserved on the reconstructed image by COSEP; the reconstructed infarct branches within the ellipse shown by COSEP are more consistent with those on the conventional high-resolution 3D MCLE image (FIG. 4) acquired after sacrifice. In FIG. 2C, COSEP presents sharper contrast between infarct and healthy myocardium than REPCOM along the signal intensity profile taken along the dashed line in FIG. 2B.We successfully demonstrated that COSEP is capable of reconstructing fine details from highly undersampled datasets. High-resolution characterization of myocardial infarction in vivo is feasible using accelerated 3D MCLE with the COSEP reconstruction. We have shown that an isotropic resolution of 1.5mm was achieved within a single breath-hold in an in-vivo prospective pig study.