Key points are not available for this paper at this time.
Abstract Arterial spin labeling (ASL) is a non‐invasive MRI technique widely used to measure cerebral blood flow (CBF), but it suffers from poor SNR, requiring the acquisition of a large number of multiple Label‐Control (L‐C) pairs at the expense of prolonged acquisition time. This work proposed a novel deep learning network to reconstruct the ASL CBF map using the minimum number of L‐C pairs. A dataset comprising 72 normal ASL raw images with 52 L‐C pairs is obtained from ADNI2. The mean CBF map, derived from 52 L‐C pairs with Gaussian smoothing and outlier cleaning, serves as the reference. The mean CBF images obtained from randomly selected 10 pairs per subject, without preprocessing, are employed as input for the proposed model. A novel data augmentation technique for model training, leveraging the characteristics of ASL L‐C pairs, is implemented. The accuracy and reproducibility of the reconstructed ASL CBF map were assessed with various state‐of‐the‐art methods on both normal and Alzheimer's Disease (AD) data. The proposed model surpasses state‐of‐the‐art methods, yielding PSNR = 23.9542, SSIM = 0.81499, and Lin's CCC = 0.9168 for the ASL CBF map with 10 L‐C pairs in normal data. For AD patients, the model achieves PSNR of 22.5932, SSIM of 0.80923, and Lin's CCC of 0.9078. The Bland–Altman plot, using the radiologist's manual GMCBF ROI, consistently aligns the estimated and reference CBF values. The novel data augmentation technique improved the ASL CBF map quantification accuracy with Lin's CCC >0.9. The proposed work is a promising approach to reconstructing the ASL CBF map using a minimum number of ASL LC pairs that can lead to low image acquisition time with a reduction in motion artifacts. The use of a novel data augmentation approach, utilizing the properties of ASL L‐C pairs, allows for the generation of a large dataset for further experimentation and analysis.
Shyna et al. (Wed,) studied this question.