Abstract
PURPOSE: Ferumoxytol-enhanced 5D free-running whole-heart cardiac magnetic resonance (CMR) provides CTA-like image quality but requires hours-long computational times when using the alternating direction method of multipliers (ADMM) method to solve the iterative reconstruction. A computationally efficient reconstruction method is needed for in-line 5D CMR reconstruction to enable routine clinical usage. METHODS: A variable projection augmented Lagrangian (VPAL) method is used to solve the spatial sparsity and temporal smoothness regularized-MRI image reconstruction optimization problem. The computation time and image quality of the VPAL reconstruction method is compared with the ADMM reconstruction method for reconstructing the 5D CMR image (x-y-z-cardiac-respiratory) in numerical simulations from functional CTA datasets and from clinical pediatric patients' raw k-space data. RESULTS: While VPAL takes more iterations to converge, computational cost of each iteration is much less than ADMM resulting in lower overall reconstruction time (10.0 ±3.7 hours using VPAL vs 18.1 ±8.1 h using ADMM, p = 1.9e-5). VPAL becomes progressively more efficient compared to ADMM when more cardiac phases are reconstructed. With the reduced reconstruction time, the image quality is not compromised, as evidenced from the qualitative assessment of image quality by radiologists. Quantitative evaluations of left ventricular ejection fraction (LVEF) and blood-myocardium border sharpness measurements show that VPAL and ADMM images reconstructions have similar image quality. CONCLUSION: Variable-Projection Augmented Lagrangian (VPAL) method achieves comparable reconstruction accuracy to ADMM while substantially reducing computational cost for 5D free-running CMR reconstruction, especially when more cardiac phases are used in reconstruction.