In ASL MRI, images acquired after spin labeling are subtracted from those acquired without spin labeling (the control condition). Due to the T1 decay of the labeled spins and the transit time, spin labeling time is usually limited to around 1-2 seconds, resulting in a low signal-to-noise ratio (SNR) [1]. Both the pairwise subtraction and the low SNR can result in signal outliers, which can significantly affect cerebral blood flow (CBF) quantification quality. We have previously shown that outliers can be removed automatically using the adaptive outlier cleaning (AOC) approach [2, 3]. In the original AOC, outliers are identified in relative to the intermediate mean. If the initial mean value is already dominated by outliers, the subsequent AOC process will favor outliers rather than exclude them. In this work, we improved the original AOC by incorporating prior knowledge about CBF images: a CBF map resembles a proton density map; CBF doesn’t vary dramatically across brain regions especially within the grey matter (GM) and white matter (WM).

Repeat 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]. After CBF calculations, both the original AOC and the priors-guided AOC (PAOC) were used to remove outliers before calculating the final mean CBF map. Mean CBF images were normalized into the MNI space using SPM12. The test-retest stability of CBF maps was then calculated using the intraclass correlation coefficient (ICC).

PAOC was implemented as shown in Fig. 1. The pseudo CBF map was created to have higher value in GM and lower value in WM and can be replaced with the M0 map after excluding CSF. The threshold for labeling timepoints as possible outliers can be adjusted as well. In practice, we found that one standard deviation is enough to label all suspicious points. Since CBF value doesn’t vary significantly within each brain tissue, similar to [7], we remove the suspicious timepoint if that results in reduced variations. Meanwhile, since CBF varies more in GM (for example, CBF is higher in the default mode network at rest), the outlier definition is further conditioned at whether a removal will result in increased GM/WM variation difference.

Fig.2 shows results from a representative subject. 2B shows that the original AOC improves CBF image quality in terms of recovering signal in the posterior part of the brain as marked by the dash green box. Fig. 2C shows that PAOC remarkably improved CBF map quality in terms of recovering signal in the green box as well as suppressing the artificial bright spot in the last slide. Fig.3 shows the results of PAOC for one representative subject’s two ASL MRI scans. By punishing both the sum and the difference of GM/WM standard deviations (std), PAOC showed more consistent image quality across different scan time (within an hour) as compared to punishing the sum of std only. Fig. 4 shows the ICC analysis results. ICC maps were thresholded at ICC>0.5, indicating a high test-retest stability across scan time. As compared to the original AOC (4A), PAOC (4B and 4C) improved ICC in both the anterior and posterior part of the brain. Punishing both the sum and difference of GM/WM std (4B) yielded better ICC (as marked by the green ovals) than punishing the sum only.We showed that incorporating knowledge about the GM WM CBF contrast and distributions into AOC improves both CBF image quality and the test-retest stability. Though the prior CBF image was created empirically, any proton density weighted image can be used to replace it. Actually, the prior image can be even removed so PAOC can be initialized by marking all timepoints as suspicious outliers, which can be then gradually verified. The performance of PAOC in clinical applications is being validated in ongoing work.