2429
Native Space Outlier Rejection (NaSOR) for Arterial Spin Labelling1Analysis, Invicro LLC, London, United Kingdom, 2MRI, Invicro LLC, London, United Kingdom
Synopsis
ASL suffers from relatively low signal-to-noise, so data cleaning strategies are required to optimise its utility. A previous method for outlier rejection of 2D-PASL data required time-consuming spatial normalization to standard space, degrading the original ASL data, and was limited to single inversion-time (TI) 2D-PASL. We therefore developed two native-space processing workflows, termed Native Space Outlier Rejection (NaSOR) and Native Space Perfusion-weighted Outlier Rejection (NaSPOR). The two native-space workflows performed comparably to an implementation of the previous standard-space method, in terms of both percentage of outliers rejected and coefficients of variation (CV) for test-retest CBF values, suggesting clinical utility.
INTRODUCTION:
Arterial spin labelling (ASL) is an MRI technique that uses magnetically labelled arterial blood as an endogenous tracer to noninvasively measure cerebral blood flow (CBF). It has demonstrated clinical value (e.g., AD1), and is currently deployed in multi-centre Core Lab (CL) studies2. Methods vary according to both labelling (e.g. PASL and PCASL) and acquisition (e.g. 2D and 3D) schemes. Due to the relatively low signal-to-noise of ASL imaging, and in particular 2D-PASL, suitable signal processing strategies are required to optimise utility and validity of acquired data (e.g. ADNI3). Furthermore, rapid assessment of data quality is required in CL studies e.g., to request a re-scan. In previous work4, a structural correlation-based outlier rejection scheme with pre-processing to remove extreme outliers (SCORE+) was presented to reject outliers in the CBF time series of 2D-PASL data. The SCORE+ algorithm used a linear affine transformation to MNI-space, which is both time-consuming and degrades the original ASL data with interpolation. Furthermore, it is restricted to single inversion-time (TI) 2D-PASL data, without arterial arrival time (AAT) estimation. To overcome these limitations, we initially develop and validate a native-space variation of the SCORE+ method, termed Native Space Outlier Rejection (NaSOR), before modifying for application directly to the native-space perfusion-weighted time series (termed NaSPOR); opening-up the possibility of extension to multi-TI PASL data such as that acquired in MINDMAPS5.METHODS:
T1-weighted volumes and single TI 2D-PASL data were acquired (Figure 1) in seven healthy volunteers (age: 34.1±9.2 years, 29% Female). Acquisition was repeated on the same day to assess repeatability. T1-weighted scans were processed using FSL’s fsl_anat6, providing bias-corrected T1-weighted images, brain masks, tissue probability maps (TPMs) and subcortical segmentations, with transformation parameters to standard (MNI) space. Automated pre-processing of PASL data was performed using FSL6: EPI series motion-corrected; mean EPI brain-extracted; M0 co-registered to mean EPI (rigid-body) and smoothed with full-width-half-maximum (FWHM) =5mm Gaussian kernel; perfusion-weighted time series computed (by pairwise subtraction of control-label images); mean EPI to T1-weighted rigid-body transformation parameters estimated (BBR), inversed, and used to transfer the anatomical segmentations to the mean EPI image in native ASL space. The co-registered, native-space TPMs were either retained unsmoothed, or smoothed (FWHM=5mm or 8mm Gaussian), and binarized with threshold=0.4 for use in NaSOR and NaSPOR (as per TPM threshold in4). The unsmoothed native-space TPMs were binarized with threshold=0.5 for region-of-interest (ROI) analysis on the mean CBF map from the various different methods. ROI analysis was also performed with subcortical segmentations.
NaSOR and NaSPOR were implemented in MATLAB7 using a combination of custom functions, MIAKAT8 and FSL6 tools. Both workflows have a built-in audit trail that records the analysis that has been performed. For NaSOR, the native-space perfusion-weighted images were converted to absolute CBF volumes using standard formula4, with parameters in Figure 2. The CBF time series was smoothed to the same extent as the TPMs. The NaSOR scheme retained the same criteria for outlier volume detection as SCORE+4. The mean CBF map was computed from remaining volumes in the native-space CBF time series. The NaSPOR scheme is similar to the NASOR scheme with the exception that the outlier detection is performed directly on the native-space, M0-normalized perfusion-weighted time series before conversion to a CBF time series. Similar to the NaSOR scheme, various smoothing options were explored with NaSPOR. For comparison with NaSOR and NaSPOR, a version of SCORE+ (termed impl-SCORE+) was implemented using a linear affine transformation to MNI-space (flirt, 12 DOF), FWHM=8mm Gaussian smoothing (of both CBF time series and TPMs), and TPM threshold=0.4 to generate segmentations for outlier rejection in MNI space.
RESULTS:
A subset of methods is reported here: NaSPOR and NaSOR (both without smoothing and with FWHM=5mm smoothing), and impl-SCORE+ as a reference method. On average 7.6±5.1% of volumes were rejected with impl-SCORE+ and there was no significant difference in this number with any of the native-space methods (Figure 3). The regional coefficients of variation (CV) for test-retest CBF values did not differ between any of the implemented methods (Figure 4), and were comparable to values reported in4. Sample mean CBF maps are presented (Figure 5).CONCLUSION:
This study provides preliminary evidence of the repeatability of two native-space 2D-PASL data cleaning methods (NaSOR and NaSPOR), suggesting similar performance to an existing standard-space method4; negating the time-consuming requirement for co-registration to MNI space, and without degradation of the original ASL data through interpolation. Importantly, NaSPOR can be applied to multi-TI ASL data, enabling data cleaning prior to joint modelling of CBF and AAT. This is currently being explored in MINDMAPS5, alongside patient-control discrimination.Acknowledgements
With thanks to the volunteers who participated in this study.References
1. Collij LE, Heeman F, Kuijer JPA, et al. Application of Machine Learning to Arterial Spin Labeling in Mild Cognitive Impairment and Alzheimer Disease. Radiology. 2016. doi:10.1148/radiol.2016152703.
2. A Randomized Study to Assess the Safety of GRF6019 Infusions in Subjects With Mild to Moderate Alzheimer's Disease. https://clinicaltrials.gov/ct2/show/NCT03520998. Posted May 10, 2018.
3. The Alzheimer’s Disease Neuroimaging Initiative (ADNI). http://adni.loni.usc.edu.
4. Dolui S, Wang Z, Shinohara RT, Wolk DA, Detre JA. Structural Correlation-based Outlier Rejection (SCORE) algorithm for arterial spin labeling time series. J Magn Reson Imaging. 2017;45(6):1786-1797. doi:10.1002/jmri.25436
5. The MIND MAPS consortium.
6. FSL. https://fsl.fmrib.ox.ac.uk/fsl/fslwiki
7. MATLAB. https://uk.mathworks.com.
8. MIAKAT. www.miakat.org
Figures
