A deep learning architecture for arterial spin labeling MRI to improve SNR with short acquisition time

Arterial spin labeling (ASL) is a completely non-invasive Magnetic resonance imaging (MRI) technique extensively utilized for quantifying Cerebral blood flow (CBF), a pivotal indicator for various neurological disorders. Accurate CBF estimation can enhance the clinical utility of ASL, but challenges remain due to low SNR, low resolution, and long scan times. Enhancing the SNR often entails acquiring a large number of ASL raw images, referred to as label and control images. However, this process can introduce various artifacts into ASL images and prolong scanning time. This study proposed a novel deep-learning framework to improve the SNR of ASL images while employing a minimal number of label and control images. The proposed Convolutional neural networks (CNNs) integrate dilated convolutions and Residual dense blocks (RDBs) to explicitly extract multiple features for reconstructing corrupted pixel points in noisy images. The experiments are conducted on ASL raw images sourced from the ADNI2 dataset, assessing results using various quantitative metrics like RMSE, PSNR, and SSIM and comparing them against state-of-the-art methods. Results show that the proposed approach enhances SNR using fewer label-control images compared to existing techniques such as U-Net, DWAN Net, Dilated Net and HUST on the ADNI2 dataset.