A Mixed Loss Joint Approach-Based Deep Learning Strategy to Enhance SNR and Resolution of Arterial Spin Labelling MRI

The arterial spin labelling (ASL) MRI is a non-invasive technique to quantify cerebral blood flow (CBF) to diagnose various neurological disorders. However, low spatial resolution, which produces partial volume effects and low SNR of ASL images, hampered CBF quantification, which can be corrected with appropriate post-processing approaches. The deep learning-oriented super-resolution technique is a promising way to enhance the overall image quality of ASL-MRI. The proposed 2D CNN architecture uses a residual dense block (RDB) as the basic building unit and a mixed loss joint strategy to increase SNR and resolution while retaining edge information. The proposed model is validated on simulated ASL-MRI data generated from structural images of ADNI in terms of different metrics such as RMSE, PSNR, and SSIM and using clinical data by two independent radiologists in terms of the visual quality control score (VQCS). Experimental results show that the proposed model outperforms state-of-the-art methods.