Accuracy and inter-subject reproducibility of default mode networks identified from ASL data
Felipe Barreto1,2, Xiufeng Li1, Amir Moheet3, Anjali Kumar3, Lynn Eberly4, Elizabeth Seaquist3, Fabrizio Esposito5, and Silvia Mangia1

1CMRR, Department of Radiology, University of Minnesota, Minneapolis, MN, United States, 2Department of Physics, University of Sao Paulo, Ribeirao Preto, Brazil, 3Department of Medicine, University of Minnesota, Minneapolis, MN, United States, 4Division of Biostatistics, University of Minnesota, Minneapolis, MN, United States, 5Department of Medicine and Surgery, University of Salerno, Baronissi, Italy

Synopsis

The present study aimed at characterizing the robustness of the default mode network (DMN) extracted at single subject level from ASL datasets with independent component analysis. Three different analyses modes were considered, including the series of perfusion weighted images, the full time series, and the pair-wise average of control/tag images (pseudo-BOLD). Results show that the three analysis modes produce DMNs with similar accuracy at a group level, but the pseudo-BOLD mode resulted in smaller inter-subject variability of the spatial distribution of the single-subject DMNs.

Purpose

Arterial Spin Labeling (ASL) methodologies are widely used to quantify cerebral blood flow in a wide range of brain conditions.1 Repeated measurements of control and tag images are required due to the intrinsic low signal to noise ratio of the technique; therefore a standard ASL acquisition requires at least 5 minutes. Taking advantage of the acquisition duration, studies have shown the feasibility of extracting resting state networks (RSNs) from such data sets.2-3 Resting state networks extracted from ASL data provide reduced yet still adequate repeatability in their spatial pattern at a group level, but their robustness in individual subjects has not been fully investigated yet. The purpose of this study was to quantify the accuracy and the inter-subject reproducibility of one particular RSN of interest, namely the default mode network (DMN), extracted at a single subject level from ASL datasets with independent component analysis (ICA). To this aim, we calculated the correlation coefficient between the z-scores maps of individual subjects and the z-score map of an external template obtained with BOLD-fMRI from a different group of healthy subjects. We also evaluated the global z-score of each subject-map within the external template. Finally, we optimized the number of IC components necessary for extracting the DMN having the highest correlation with the external template in each subject.

Methods

ASL datasets were obtained within the context of a CBF study which recruited so far 13 individuals with type 1 diabetes (6M/7F, 34.8±10.9 years). Data were acquired using a 3T Siemens Prisma system and a pseudo-continuous ASL sequence. Parameters were: TE/TR=14/5000 ms, labeling time/post labeling delay=1600/1600 ms, matrix of 70x70, FOV=210 mm, slice thickness=3 mm, 36 slices, gap of 3.6 mm and 160 volumes. Subjects stayed awake with eyes closed during the experimental session. We used BrainVoyagerQX (Brain Innovation, Maastricht, The Netherlands) for slice scan time correction, motion correction, and normalization to Talairach space. Analyses were performed on three different modes: perfusion weighted images, dM (pairwise subtraction of control and tag images); full time series, FS; pseudo-BOLD images (pair-wise average of control and tag images). Datasets were high-pass filtered with a cut-off of 0.008 Hz, and spatially smoothed with a Gaussian kernel of 6 mm. Multiple ICA runs were performed using ICASSO 4 with 10 repetitions in correspondence a variable number of independent components (IC) ranging from 20 to 40 in steps of four. The external DMN template served as input for a custom-built script that selected the most similar IC map to the template based on the correlation coefficient between voxel z-scores of each map and the template. The selected maps were visually inspected and z-scores were extracted using a mask created from the external template thresholded at z=2. Statistic summaries were finally obtained for each of dM, FS and pseudo-BOLD modes.

Results and discussion

Visual inspection of the DMNs extracted by ICA confirmed good spatial similarity with the external template in 12, 10 and 9 out of 13 subjects for the pseudo-BOLD, FS and dM modes, respectively. At a group level, the three analysis modes produced highly similar DMN maps (Figure 1) and almost identical mean z-scores extracted from the external DMN mask (Table 1). However, FS and pseudo-BOLD resulted in better inter-subject reproducibility, with inter-subject deviations approximately 40% smaller than those obtained with dM, potentially caused by the low pass filtering effect induced by the 2-point averaging. The mean correlation coefficients revealed that in average the accuracy of the individual DMNs obtained with the three analysis modes were also very similar, but pseudo-BOLD produced again smaller inter-subject deviations. The optimal number of ICs necessary to produce the most robust DMN networks was slightly larger for the FS mode as compared to the other analysis modes, possibly due to the extra variability included by the labeling paradigm.

Conclusion

ASL datasets can be effectively used to extract accurate DMNs on an individual subject basis. Among the three analysis modes evaluated in this study, the pseudo-BOLD mode resulted in smaller inter-subject variability of the spatial distribution of the DMNs.

Acknowledgements

NIH grants: P41 EB015894, P30 NS076408, R01DK099137.

References

[1] Alsop et al. Neuroimage 2015;73:102 [2] Zhu et al. PLoS One 2013;8:e65884 [3] Jann et al. NeuroImage 2015;106:111 [4] Himberg et al. NeuroImage 2004;22:1214.

Figures

Figure 1. Group-based DMN maps (n=13) calculated for the three different analysis modes.


Table 1. Average z-scores extracted from the DMN maps using the mask of the external template, mean correlation coefficient within the external template and optimal number of ICs for the three analysis modes.



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
1658