Evaluation of Deep Learning Reconstruction in 4D-pCASL-based MR Angiography using CENTRA-Keyhole and View-sharing

We evaluated the effectiveness of deep learning (DL)-based reconstruction in improving image quality of 4D-pCASL-based angiography with CENTRA-Keyhole and View-sharing (4D-PACK) compared with Compressed Sensing (CS). Ten healthy volunteers underwent 4D-PACK imaging on a 3.0T MRI scanner at reduction factors of 6, 8, 10, and 12 using both reconstructions. Quantitative metrics included contrast-to-noise ratio (CNR), signal-to-noise ratio (SNR), and contrast ratio (CR) in M1–M4 segments of the middle cerebral artery. Temporal profiles were assessed via time–intensity curves, and three experienced technologists performed qualitative evaluations. DL achieved significantly higher CNR and SNR across all factors (p < 0.01), with mixed CR results favoring DL at lower factors. No significant differences were observed in temporal signal profiles. Visual scores were significantly higher for DL, particularly at factor 6, and remained high even at greater accelerations. DL enables substantial reduction in scan time while preserving vascular detail, supporting its clinical utility for non-contrast-enhanced 4D-MRA.