Introduction

3D Spiral TSE and GraSE techniques are often recommended for brain Arterial Spin Labeled (ASL) MRI due to their high Signal-to-Noise Ratio (SNR) and compatibility with optimized background suppression [1]. However, these methods can be challenging in regions with increased B0 inhomogeneities (such as the skull base), potentially leading to signal loss and geometric distortion. An alternative approach, 3D Cartesian TSE, shows promise for ASL and has demonstrated increased robustness to B0 inhomogeneities in both brain and kidney ASL applications [2-4]. Nevertheless, due to its inherently lower SNR, brain ASL images are often acquired at reduced spatial resolutions (e.g., 3.5x3.5x6mm³), prompting a desire for higher resolution ASL imaging within clinically-feasible scan times. Recently, a compressed sensing artificial intelligence (CS-AI) framework, leveraging an adaptive-CS-Network, has emerged as a potential solution to accelerate various MR imaging techniques, demonstrating promise across different applications [5]. Consequently, this study aims to assess the viability of utilizing CS-AI to accelerate brain ASL imaging using 3D Cartesian TSE with higher spatial resolution.

Methods

Pseudo-continuous ASL (pCASL) with background suppression (BGS) was used for brain perfusion imaging on a 3T Philips scanner under IRB approval. (Ingenia Elition, Philips Healthcare, Best, The Netherlands). Five healthy volunteers were scanned in an axial orientation. Vendor-supplied 3D TSE Cartesian acquisition with low-high profile ordering and a radial turbo direction was used. Other parameters were: TR/TE = 4700/12 ms, FOV = 240x240x96 mm³, matrix = 96x96 with 32 slices, acquired resolution = 2.5x2.5x6 mm³, reconstructed resolution = 2.15x2.15x3 mm³, ETL = 80, echo spacing = 2.1 ms, label duration = 1.8 s, post-label delay = 1.8 s, NSA = 1, and 4 BGS pulses. CS-SENSE acquisitions with 2-fold, 3-fold, and 4-fold accelerations were performed in 2:50 min, 1:53 min, and 1:34 min respectively, and the acquired data were subsequently reconstructed with a vendor-supplied artificial intelligence (AI)-based algorithm. Additionally, for comparison, SENSE acquisitions, with 2-fold, 3-fold, and 4-fold accelerations were also performed in 2:22 min, 1:53min, and 1:25 min respectively. In addition, ASL acquisition with no acceleration was performed in 4:25 min, serving as a benchmark for image quality. For CBF quantification, M0 images were acquired using CS-AI with 2-fold, 3-fold and 4-fold accelerations in 1:25 min, 0:56 min, and 0:47 min respectively, as well as with no acceleration in 2:12 min.
CBF quantification was performed using the recommended approach [1], CBF maps were subsequently generated for each of the ASL/M0 acquisitions with CS-AI, and for the acquisitions with no acceleration. Circular ROIs were manually drawn on gray matter, and the mean CBF values within the ROIs were extracted. The ROIs covered a variety of brain regions, including upper frontal, upper parietal, lower operculum, lower occipital, thalamus, cerebellum, and brainstem. Linear regression and Bland-Altman analysis were used to compare the agreement of the ROI-based mean CBF values calculated from the ASL/M0 images acquired with 2-fold accelerated CS-AI, and with no acceleration.

Results

Brain perfusion weighted images acquired with improved spatial resolution (2.5x2.5x6 mm³) using CS-AI showed good SNR (fig.1). With 2-fold (fig. 1a) and 3-fold (fig. 1b) accelerations, the images acquired using CS-AI showed significantly higher SNR and better robustness to signal loss than the corresponding images acquired using SENSE with the same accelerations. CBF maps calculated from the ASL/M0 images acquired using CS-AI with 2-fold, 3-fold, and 4-fold accelerations are similar to the CBF map acquired with no acceleration (fig. 2a). In addition, the CBF distributions of CS-AI with different accelerations are consistent with the CBF map acquired with no acceleration (fig. 2b). This indicates the use of CS-AI with up to 4-fold acceleration does not introduce significant bias to the CBF quantification. The CBF values acquired using no acceleration and using CS-AI with 2-fold acceleration showed high correlation (R² = 0.91, fig. 3a), as well as excellent agreement without significant bias (fig. 3b).

Conclusion

We applied a compressed sensing artificial intelligence (CS-AI) framework to accelerate the brain ASL MR Imaging using 3D Cartesian TSE with higher spatial resolution (2.5x2.5x6 mm³). Perfusion weighted images acquired using CS-AI showed good SNR and image quality, and the SNR is significantly higher than the corresponding images acquired with SENSE. In addition, the use of CS-AI does not introduce bias to CBF quantification. CS-AI enables higher resolution 3D ASL MRI in the brain in a clinically feasible scan time (ASL/M0 acquisition in less than 4 min with 2-fold acceleration). This could be particularly useful to measure perfusion alterations in small-sized brain lesions, such as small metastatic and ischemic lesions.

Acknowledgements

No acknowledgement found.