Hu Cheng^{1}, Ao Li^{1}, Andrea Avena-Koenigsberger^{1}, Chunfeng Huang^{1}, and Sharlene Newman^{1}

As an alternative to template based brain parcellation in functional connectivity analysis, nearly equal-sized random parcellations are applied to individual subjects multiple times to obtain a pseudo-bootstrap sample of the functional network. As one application, the method was applied on the HCP resting state dataset to identify individuals across scan sessions based on the mean functional connectivity. With a parcellation number of 278 and bootstrap sample size of 400, an accuracy rate of ~90% was achieved by simply finding the maximum correlation of mean functional connectivity of pseudo-bootstrap samples between two scan sessions.

Resting state functional data from 87 subjects were downloaded from the data release of the HCP (https://www.humanconnectome.org/ ) (Q1 through Q3). Each subject has two sessions of rsfMRI scans: REST1 and REST2, which is one day apart. The dataset have been preprocessed and normalized the MNI template via nonlinear transformation. 400 random parcellation with 278 roughly equal-size parcels on the gray matter of the MNI template was generated using an algorithm by growing voxel neighborhoods around a set of randomly selected voxel-seeds based on geodesic distance. Functional connectivity was computed as the Pearson pair-wise correlation between the time series of the nodes after regressing motion parameters as well as signal from the white matter and CSF, resulting a 278 ´ 278 matrix for each parcellation. Each subject has 400 such FC matrices per session, and the mean of the FC (mFC) forms a vector of 400 elements. This vector was named as the mean FC vector (mFCV). To use the FC as a fingerprint to identify subjects across resting state fMRI scans3, the cross-correlation of the mFCV between all subjects in REST1 and all subjects in REST2 was calculated. A subject j in REST2 was predicted to be subject i in REST1 if the correlation $$$C_{ij}$$$ between node j and i is the largest among all subjects in REST2, i.e.

*$$C_{ij}>C_{ik}, for \space k\neq j$$.*

The same principle applied to matching a subject in REST1 a target subject in REST2. The prediction accuracy rate (PCR) was defined as the number of correct matching normalized by the total number of subjects. The effect of number of PBS samples on the PCR was assessed by reducing the number of random parcellation from 400 to 110. As a comparison, we also evaluated the PCR using the fingerprinting method proposed by Finn et al., which was based on the cross-correlation of the FC matrix derived from one single parcellation. The PCR was calculated 40 times with different random parcellation out of the 400.

**Results**

**Discussion**

Figure
1. Each pseudo-bootstrap random parcellation gives rise to a FC matrix and a
mean FC, an example of the distribution of the mFC from 400 samples is shown in
(A). The mFC values of all subjects in REST1 and REST2 along with their
differences are plot in (B).

Figure
2. Correlation matrix of the inter-subject mFC derived from the Pearson
pair-wise cross-correlation between the vectors of 400 mFC from
Pseudo-Bootstrap samples. This correlation matrix represents the likelihood
between subjects in terms of the coherence of change of mFC with
parcellation.

Figure
3. The prediction accuracy rate (PCR) as a function of number of
pseudo-bootstrap samples.

Figure
4. FC fingerprinting prediction accuracy rate based on the correlation of FC
from template based parcellation. The template was drawn from the PBS
samples.