Functional brain activity has been widely assessed with the blood-oxygen-level-dependent (BOLD) fMRI. But BOLD signal is relative and contaminated by neurovascular effects. It can even be offset the target brain area since the major contribution of BOLD signal change comes from the oxygen level change in vein. These issues can be greatly avoided by using arterial spin labeling perfusion fMRI, which directly measures the quantitative cerebral blood flow (CBF). However, ASL MRI has a low signal-to-noise ratio (SNR) [1], which has greatly hindered a widespread application of ASL in dynamic functional activity studies. Robust principal component analysis (RPCA) [2] is a new technique for feature extraction or data recovering. Sorting all data into a big data matrix, RPCA decompose split it into a low-rank (L) plus a sparse (S) component. The L-component represents the slowly varying part and the S-component carries the temporally uncorrelated dynamic changes. Such process fits well with ASL signal denoising since ASL signal can be roughly modeled as a combination of background CBF and random noise plus other physiological noise [3]. The purpose of this study was to evaluate RPCA (or equivalently L+S decomposition) in ASL fMRI.ASL data acquired with a pseudo continuous ASL sequence were identified from previous study [4] and local database. Data acquisitions were approved by IRB with signed consent form from all subjects. 15 old subjects were included. ASL images were preprocessed using ASLtbx with state-of-art denoising procedures [5, 6]. Each CBF image volume was then reformatted into a column vector, and all CBF column vectors were then sorted into a big matrix M (Fig. 1). The inexact augmented Lagrange multiplier RPCA algorithm [7] and the code provide by Chen and Ganesh were used to decompose M into the L and S component. The weight on sparse error term in the cost function was set to be 1/sqrt(number of inner brain voxels). The bottom row of Fig. 1 shows an example of L+S decomposition for one slide of one timepoint. After decomposition, each L column vector was reshaped back to a 3D L-CBF image. Both the original CBF images and the L-CBF images were spatially warped into the MNI space using SPM12. Seed-based functional connectivity (FC) analyses were performed. The seeds were two regions-of-interest (ROIs) drawn in right motor cortex (MC) and parietal cortex (PA), respectively (Fig. 2). The mean CBF values were then extracted from each ROI from all CBF images. The time series of the mean CBF value was then used as the regressor for a whole brain simple regression analysis. The correlation coefficient maps of all subjects were grouped together for a group level statistical analysis. The same analysis was performed for the original and the L-CBF images separately.Fig. 1 shows that the L+M decomposition can extract a good quality baseline CBF map even from one timepoint. Fig. 3 shows the average CBF map of the M, L, and S CBF image series. As compared to the original mean CBF map, the mean L-CBF map presents higher homogeneity and roughly while keeps the same grey matter/white matter image contrast. Fig. 4 shows the results of FC analysis. The top row was the group level MC-FC analysis results; the bottom row was PA-FC results. On the left was based on the original CBF images; on the right was from the L-CBF images. Significance level was defined by a Bonferroni corrected p=0.05. For both FC, L+S decomposition greatly improved sensitivity for identifying the contralateral FC. For MC-FC, L+M showed FC in the middle superior MC as well as in the supplementary motor area. For PA-FC, L+M showed markedly increased FC in precuneus and several other brain areas. We showed that L+M decomposition can extract reasonable CBF map from even one time-point. The mean CBF map after taking the sparse components out presents reduced inhomogeneity though the global CBF value also reduced, which may be improved by adjusting the L+S algorithm parameters though a test-retest study or a simulation study would be needed to find the optimal parameter values. The most significant improvement of L+M for ASL is in the time-series analysis, where we showed a remarkably increased sensitivity for two different seeds-based FC. L+M decomposition may eventually pave the way of ASL MRI for time series analysis such as FC, functional connectome analysis, resting state network analysis etc.