Synopsis

Keywords: Data Acquisition, Diffusion Tensor Imaging, High resolution 3D DTI

To propose an accelerated high resolution whole brain DTI and investigate its feasibility (1.0$$$mm^3$$$ isotropic spatial resolution, imaging time ~ 15 min)

Introduction

Diffusion tensor imaging (DTI)¹ is to provide information on anisotropic diffusivity and hence characterize microscopic architecture. Nevertheless, it is still challenging to achieve highly well-defined diffusion information in whole brain MRI due to limited spatial resolution and prolonged imaging time. Whole brain DTI with single-slab 3D imaging is promising method with high signal-to-noise ratio (SNR) and no slab boundary artifact, but recent studies have some limitations in terms of the low spatial resolution (>1 $$$mm^3$$$)², insufficient whole brain coverage, and low magnetic field strength (<3T)³. To address these issues, we first propose accelerated diffusion weighted imaging using single-slab 3D segmented EPI with pseudo-linear random sampling in k-q space to be capable of reduced scan time. Second, we combine locally low rank constrained reconstruction to achieve artifact-free diffusion weighted images by exploiting spatial-angular redundancies. Finally, we investigated the feasibility of accelerated high resolution whole brain DTI on a 3T scanner (1.0$$$mm^3$$$ isotropic spatial resolution, imaging time ~ 15 min).

Methods

High resolution single-slab 3D EPI
To achieve our aim, a single-slab 3D segmented EPI with pseudo-linear random sampling in k-q space was developed. Fig 1 represents a schematic of the proposed single-slab 3D segmented EPI (diffusion weighted imaging and phase navigator to correct shot-to-shot variations) (Fig. 1a) and its corresponding incoherent encoding schemes in k-q space (Figs. 1b, c). Especially, we use a grouped, pseudo-linear undersampling strategy with some physical constraints which is employed to achieve incoherence and alleviate both amplitude and phase modulation which occur considerably along the 3D EPI readout simultaneously. With this grouping sampling schedule, the blip gradient amplitudes in the ky and kz directions are physically constrained to achieve smooth signal transition while avoiding abrupt gradient switching. Therefore, variable density incoherent undersampled k-space can be acquired without significant k-space modulation and is utilized for robust constrained reconstruction.

Spatial-angular low rank constrained reconstruction
Missing signals were interpolated by solving constrained optimization problem with spatial and angular, locally low rank priors exploiting the fact that the different diffusion contrasts share the brain structure in images. The formulation of the problem with additional sparsity prior is:
$$\mathcal{J} \left \{ \mathrm{\mathbf{X}=\left( \mathbf{x}_0,...,\mathbf{x}_{N_d-1}\right)} \right \} = \lambda_p\sum_{p=0}^{N_p-1}\left\|\mathcal{R}_p\left ( \mathbf{X} \right ) \right\|_{*} + \lambda_{\Psi}\sum_{d=0}^{N_d-1}\left\|\Psi\left ( \mathbf{x_d} \right ) \right\|_{1} \\s.t. \quad \mathbf{Y} = \mathbf{F_u}\mathbf{S}\mathbf{P}\mathbf{X}$$
where $$$\mathbf{Y}$$$ denotes the acquired k-space signal; $$$\mathbf{P}$$$ is the phase map for each shot and each diffusion direction; $$$\mathbf{S}$$$ is the sensitivity encoding operator; $$$\mathbf{F_u}$$$ is the Fourier transform operator with undersampling; $$$\mathbf{X}$$$ is the desired artifact-free DWI, which is $$$\mathbf{X} = \mathbf{vect}([\mathbf{x}_0,...,\mathbf{x}_{N_d-1}])$$$; $$$N_d$$$ is the number of diffusion encoding directions; $$$N_s$$$ is the number of shots in each DWI; $$$N_c$$$ is the number of coils. This problem is solved using variable splitting methods under the framework of alternating direction approach.

Experiments
Experimental studies were performed in 15 healthy volunteers on a 3T whole-body MR scanner (Prisma, Siemens Healthineers, Erlangen, Germany) using the proposed, rapid diffusion-weighted MR pulse sequence with a 52-channel head coil. The imaging parameters were: TE = 64ms, field-of-view (FOV) = 230×200×150 $$$mm^3$$$, matrix size = 220×192×144, readout bandwidth = 1140 Hz/Pixel, echo train length (ETL) = 32, and Flip Angle =90◦. Electrocardiogram synchronization was used to minimize periodic phase incoherence from cardiac motion. Six diffusion-directions with b=1000s/$$$mm^2$$$ were acquired for DTI study. In the prospective study, the reduction factor is 5.5 (total scan time: 15min). Additionally, the 2D single-shot EPI with 1×1×2mm resolution was acquired for comparison with the proposed method. Fractional anisotropy maps were generated using FSL⁴.

Results

Fig 2 shows the brain images reconstructed using a conventional SENSE (without priors) and proposed method in the absence (‘before PC’) and presence (‘after PC’) of inter-shot 3D phase correction. The images with different diffusion encoding directions were shown in different rows with SSIM. The proposed method with inter-shot 3D phase correction reveals higher SSIM with noise suppression. Fig 3 shows the repeated comparison with prospective undersampled data (R =5.5, total scan time ~ 15min). DWIs, FA maps, and colored FA maps are shown in different rows. White arrows indicate that the proposed method can delineate the fine structures in DTI without noise amplification. Finally, we report the comparison of whole brain DTI colored FA maps acquired using a conventional 2D single-shot EPI and the proposed method. The zoom-in images show the details of the brain structure, and the proposed method is superior in preserving fine details.

Discussion and Conclusion

We successfully demonstrated that the accelerated, 3D diffusion tensor imaging can be produced over the whole brain roughly in 15 minutes without an apparent artifact and amplified noise. We need to further investigate It is expected that the proposed method widens the neurological and clinical utility of diffusion tensor imaging, although it needs to be further investigated.

Acknowledgements

This work is supported in part by NRF-2018M3C7A1056887, KMDF-202011B35, and KMDF-202011C20.