Whole-Brain Multi-Parameter mapping using Dictionary learning

Targeted Audience: Researchers and Clinicians interested in using quantitative relaxation measures to the study the effect of physiological and metabolic changes including those associated with neurological and psychiatric disorders on tissue parameters. Purpose: Quantification of multiple tissue parameters is emerging as a powerful tool in diagnosing various neurological and psychiatric diseases. However the major bottleneck in its routine clinical use is the long acquisition times needed to acquire multiple contrast weighted images. In addition, long acquisition times are likely to result in motion induced artifacts. In this work, we propose a dictionary learning based scheme to simultaneously recover T1ρ and T2 maps. Although, T1 ρ is sensitive to a number of tissue properties of interest, it is not specific. Acquiring additional parameters (e.g. T2ρ and T2) can improve the specificity of T1ρ. The proposed method models the data as a weighted linear combination of basis functions from a dictionary, which is learned from the measured data. Methods: The reconstruction from under-sampled data is posed as a constrained optimization problem given by: , , ; 1, where b=under-sampled data, U=sparse coefficient matrix, V=learned dictionary, and A considers coil sensitivity encoding along with Fourier encoding. A sparsity-promoting l1 norm prior is enforced on U, while the Frobenius norm of the dictionary V is constrained. This approach jointly estimates the sparse coefficients U and the dictionary basis functions V from the measured undersampled k-space data b. We evaluated the utility of the proposed algorithm using both retrospective undersampling and prospective undersampled acquisitions. A fully sampled 2D dataset was acquired using a TSE sequence combined with T1ρ and T2 preparatory pulses (Turbo factor=8; FOV=22x22cm, TR=2.5s, spin lock freq=330Hz). T1ρ and T2 weighted images were acquired by changing the duration of the T1ρ preparation pulse spin lock time (TSL) and T2 preparation pulse echo time (TE). The data was collected for 12 equispaced TSLs and TEs each ranging from 10ms to 120ms resulting in a scan time of 16min. The 2D data was retrospectively undersampled using a variable density sampling pattern at acceleration R=6,8,10,12. The 3D prospective dataset with R=8 was acquired using a segmented 3D GRE sequence based on the 3D MAPSS approach (TR=5.6ms, Res=1.7mm isotropic, FOV=22x22x22cm). Ten T1ρ spin-lock images and ten T2 TE images were acquired (10ms 100ms) resulting in a scan time of 20min. Both the datasets were reconstructed using the proposed scheme and the results were compared with kt-PCA and CS based schemes. The proposed scheme was solved using two algorithms: Algorithm 1without variable splitting and Algorithm 2 – with variable splitting. The parameters were estimated using a single