Accelerating multi-echo brain MRI at 7T using variable-density sampling & patch-based regularization

1877 

Abstract Submission 

Jyoti Mangal¹, Donovan Tripp¹, Rene Botnar^1,2,3, Claudia Prieto^1,2,3, David Carmichael¹

¹School of Biomedical Engineering and Imaging Sciences, King's College London, London, United Kingdom, ²School of Engineering, Pontificia Universidad Católica de Chile, Santiago, Chile, ³Institute for Biological and Medical Engineering, Pontificia Universidad Católica de Chile, Santiago, Chile

Jyoti Mangal  
School of Biomedical Engineering and Imaging Sciences, King's College London
London, United Kingdom

Donovan Tripp  
School of Biomedical Engineering and Imaging Sciences, King's College London
London, United Kingdom

Rene Botnar  
School of Biomedical Engineering and Imaging Sciences, King's College London|School of Engineering, Pontificia Universidad Católica de Chile|Institute for Biological and Medical Engineering, Pontificia Universidad Católica de Chile
London, United Kingdom|Santiago, Chile|Santiago, Chile

Claudia Prieto  
School of Biomedical Engineering and Imaging Sciences, King's College London|School of Engineering, Pontificia Universidad Católica de Chile|Institute for Biological and Medical Engineering, Pontificia Universidad Católica de Chile
London, United Kingdom|Santiago, Chile|Santiago, Chile

David Carmichael  
School of Biomedical Engineering and Imaging Sciences, King's College London
London, United Kingdom

Quantitative mapping is beneficial for a range of neurological applications such as identification and characterisation of small-scale brain architecture[1] but scan times are often long. We previously showed that optimal precision of T2* estimation from multi-echo gradient echo (MEGRE) sequences is achieved at repetition times (TRs) of >30ms at 7T[2]. Even accelerated high-resolution (<0.6mm) multi-parametric maps using parallel imaging[3] with these parameters have clinically infeasible durations and/or remaining aliasing and artefacts at large acceleration factors. Here, we investigated the potential of a patch-based low-rank reconstruction method (HD-PROST)[4] with a variable-density (VD CASPR)[5] undersampling of k-space to efficiently accelerate MEGRE data. We assess the image quality of the reconstructed echo volumes visually and using root mean squared error values.

Data acquisition: Data acquisition was performed at 7T (Siemens TERRA) for two healthy volunteers (HV). For HV1, a 10-echo GRE at 1mm³ resolution dataset was acquired with the total time of acquisition Tacq =12:26[min:s]. and for HV2, a 7-echo GRE at 0.6mm³ was acquired with Tacq=24:50[min:s]. Both data was fully sampled with an elliptical shutter. The detailed sequence parameters are given in Fig.1 and 2.

'Ground Truth' Reconstruction: The fully sampled data was reconstructed with CG-SENSE[6] for each echo (tolerance=10^-10, maxiter=15).

Undersampling with VD-CASPR was used in the phase-encoding plane following spiral-like interleaves on the Cartesian grid with denser sampling of k-space centre while maintaining a golden angle step [7] between consecutive spirals. For acceleration factors 6, 8 & 10, two sampling scenarios were considered (a) using identical sampling trajectories for all echo volumes, and (b) sampling the echo volumes such that the spiral interleaves between subsequent volumes were rotated (shifted) to introduce aliasing incoherence.

HD-PROST reconstruction: HD-PROST is a regularization framework that performs an iterative low-rank high-order singular value decomposition (HOSVD) of tensors obtained from multiple contrasts (with structural and contrast similarity)[5] which in our case were the multiple echo volumes with different T2* weighting. The low-rank thresholding parameter σ was empirically set to 0.0051.

CG-SENSE reconstruction was also performed on the undersampled data for comparison to HD-PROST. Coil compression8 to 16 channels was performed prior all reconstructions. Sensitivity maps for were calculated from the k-space centre using ESPIRiT[8].

Analysis: Visual comparison was made between the undersampling scenarios (a) and (b) for acceleration factor 6, 8 & 10. Root Mean Square Error (RMSE) within the brain (segmented using spm12) quantified differences between each reconstruction and the ground truth (fully sampled CG-SENSE) reconstruction.

Fig 1 shows the results of the reconstructions for HV1 and HV2 for acceleration factors 1,6,8 and 10. The HD-PROST and HD-PROST* reconstructions correspond to the sampling scenario (a) and (b) respectively. Fig. 2 shows the whole brain average RMSE for both HVs. RMSE is reduced for HD-PROST* reconstruction results compared to HD-PROST for all cases showing a clear advantage to altering the phase-encoding between echoes to decrease artifact coherence across the echo-train. Comparison of performance with compressed sensing[9] was not explored but previous work[5] showed HD-PROST was superior.

HD-PROST enables significantly greater acceleration potential than standard parallel imaging approaches owing to the exploitation of information redundancy across contrasts consistent with previous work. We have used redundancy within ME-GRE as the multi-echo images are inherently coregistered and display similar contrast making them suitable for this reconstruction approach.

Methods Development ¹

Anatomical MRI ²

Acquisition

HIGH FIELD MR

MRI PHYSICS

STRUCTURAL MRI

Other - Quantitative MRI; Reconstruction; Image acceleration;