3700
Identification and Characterization of Resting State Networks in the Translational Pig Model1Physics and Astronomy, University of Georgia, Athens, GA, United States, 2Bio-Imaging Research Center, University of Georgia, Athens, GA, United States, 3University of Georgia, Athens, GA, United States, 4Foods and Nutrition, University of Georgia, Athens, GA, United States, 5Animal and Dairy Science, University of Georgia, Athens, GA, United States
Synopsis
Due to the similar size, structure, composition, and neurodevelopment of the pig brain in comparison to the human brain, the pig serves as a valuable large animal model for studying brain connectivity. Presented here are five resting state networks (RSNs) identified within the three-week-old piglet brain determined using temporal sparse dictionary learning. Each RSN’s learned activation map correlates well with a constructed pig RSN atlas with Pearson spatial correlation coefficients in the range of [0.30 0.53], and clear differences in the power spectra, as well as unique characteristic frequencies, associated with the learned signal for each RSN are observed.
Purpose
To identify and characterize resting state functional connectivity networks (RSNs) in the translational pig model.Methods
Using a GE Signa HDx 3T scanner and a HD quadrature knee coil, resting state fMRI (rs-fMRI) and T1-weighted anatomical data were acquired for three-week-old Landrace-cross piglets (n=12) using the following two sequences: 1.) 3D GE EPI (TR=3s, TE=30ms, FA=80°, FOV=12.8x12.8x6.4cm, a matrix size of 96x96x32, 300 total volumes, and an acquisition time of 15 mins) and 2.) 3D FSPGE (TR=5.5s, TE=2.1ms, FA=9°, FOV=12.8x12.8x6.4cm, slice thickness=1mm, and a matrix size of 256x256x112), respectively. Piglets were anesthetized and maintained with 1.5% inhalational isoflurane in oxygen.
Each piglets’ rs-fMRI data was preprocessed using the Statistical Parametric Mapping (SPM) software1 to realign images to correct for motion, perform slice-timing correction, and perform spatial normalization. One piglet was chosen as the template, and the rs-fMRI datasets of the other eleven piglets were spatially normalized to the template piglets’ fMRI space. Brain tissue was then separated from the skull and other surrounding tissues. Next, the rs-fMRI time-series from the twelve piglets were temporally concatenated into a group dataset. Activation maps were determined from this group dataset using four decomposition methods: spatial independent component analysis (ICA), temporal ICA, spatial sparse dictionary learning (DL), and temporal DL. ICA was performed using the Group ICA of fMRI Toolbox (GIFT),2 and the data was decomposed into 20 independent components. DL was performed using the SPArse Modeling Software (SPAMS) toolbox,3 and the data was decomposed into 100 atoms or components.
The activation maps generated by each of the four methods and a standard pig brain atlas4 were then spatially normalized to the anatomical space of the template piglet. In order to investigate how closely pig RSNs resemble human RSNs, a pig RSN atlas was created by using anatomies that are known to be associated with certain human RSNs.5 This pig RSN atlas contained the following five RSNs: executive control (EX), cerebellar (CERE), visual (VIS), sensorimotor (SM), and auditory (AUD).
To determine which method best identifies these five RSNs, the activation maps from each method that produced the maximum Pearson spatial correlation coefficients with each individual RSN atlas were compared, and to further characterize each RSN, the power spectra associated with the learned signal components of these maps were also examined. Each power spectrum was divided into three frequency segments for evaluation: 0.01-0.027 Hz (Slow-5), 0.027-0.073 Hz (Slow-4), and 0.073-0.198 Hz (Slow-3).6
Results
The maximum Pearson values for each method and RSN are given in Figure 1. Temporal DL tends to produce activation maps that correlate better with the created pig RSN atlas than the other three methods. Since temporal DL produced multiple maps with similarly high Pearson values, averaged temporal DL activation maps were created by averaging the top three maps with the highest Pearson values for each RSN. These averaged maps produce even stronger correlations with each RSN from the atlas (Fig. 1). Figure 2 displays representative cross sectional images of the averaged activation maps and created pig RSN atlas for each RSN, all overlaid on the corresponding anatomical images. Clear differences in the frequency distributions in the power spectra associated with the signal components learned by temporal DL for each RSN are also observed (Fig. 3), and the dominant frequencies associated with each RSN can be seen more clearly by examining the smoothed power spectra (Fig. 4).Discussion and Conclusion
Although temporal DL tends to produce activation maps that better correlate with the created pig RSN atlas, this does not necessarily mean that temporal DL is superior to the other methods in all cases. One limitation of this work is the small sample size (n=12). For larger sample sizes, the other methods examined in this work may possibly produce improved results. However, in this work, temporal DL produces Pearson values in the range of [0.30 0.53], which falls within the typical range reported in the literature.5,7,8 By averaging the top three temporal DL maps with the highest Pearson values, an averaged map can be generated that produces even stronger Pearson correlations. These maps demonstrate five well-defined RSNs within the piglet brain. Another limitation of this work arises when characterizing each of these RSNs based on its power spectrum. Since a TR of 3s was used during data acquisition, the maximal frequency examinable is 0.167 Hz. This value lies in the middle of the frequency range defined as Slow-3, and higher frequencies that lie within the Slow-2 and Slow-1 ranges are unobservable.6 However, clear differences in the power spectra are observed, and unique characteristic frequencies associated with each RSN can be identified.Acknowledgements
This work was supported in part by NIH RO1 grant RO1NS093314 and UGA faculty research grant.References
1. Penny WD, Friston KJ, Ashburner JT, Kiebel SJ, Nichols TE. Statistical parametric mapping: the analysis of functional brain images. Elsevier; 2011.
2. Calhoun V, Adali T. Group ICA of fMRI toolbox (GIFT). Online at http://icatb. sourceforge. net 2004.
3. Mairal J, Bach F, Ponce J, Sapiro G. Online dictionary learning for sparse coding. 2009. ACM. p 689-696.
4. Saikali S, Meurice P, Sauleau P, Eliat P-A, Bellaud P, Randuineau G, Vérin M, Malbert C-H. A three-dimensional digital segmented and deformable brain atlas of the domestic pig. Journal of neuroscience methods 2010;192(1):102-109.
5. Smith SM, Fox PT, Miller KL, Glahn DC, Fox PM, Mackay CE, Filippini N, Watkins KE, Toro R, Laird AR. Correspondence of the brain's functional architecture during activation and rest. Proceedings of the National Academy of Sciences 2009;106(31):13040-13045.
6. Gohel SR, Biswal BB. Functional integration between brain regions at rest occurs in multiple-frequency bands. Brain connectivity 2015;5(1):23-34.
7. Lois G, Linke J, Wessa M. Altered functional connectivity between emotional and cognitive resting state networks in euthymic bipolar I disorder patients. PLoS One 2014;9(10):e107829.
8. Brookes MJ, Woolrich M, Luckhoo H, Price D, Hale JR, Stephenson MC, Barnes GR, Smith SM, Morris PG. Investigating the electrophysiological basis of resting state networks using magnetoencephalography. Proceedings of the National Academy of Sciences 2011:201112685.
Figures
