Single Scan, Subject-Specific component extraction in dynamic functional connectivity using COBE.
Poster No:
1729
Submission Type:
Abstract Submission
Authors:
Pratik Jain1, Anil Sao2, Bharat Biswal3
Institutions:
1New Jersey Institute of technology, Newark, NJ, 2Indian institute of Technology Bhilai, Bhilai, Chhattisgarh, 3New Jersey Institute of Technology, Newark, NJ
First Author:
Co-Author(s):
Introduction:
The study of individual differences in healthy controls can provide precise descriptions of individual brain activity. Subject-specific (SS) information can be utilized to map individual brains to individual behaviors which can elucidate the mechanistic understanding of how neural events give rise to cognition and action [2]. However, the use of FC at the individual level has been a challenge due to the heterogeneity in the data acquired multiple times from the same subject (within subject variation). In our previous work, we attempted to extract the SS components using multiple scans of the same subject [3]. However, acquiring multiple scans of a subject is difficult, hence in this work we propose to extract the SS components using a single scan of the subject.
Methods:
Publicly available Human connectome project (HCP) dataset consisting of 100 unrelated subject's (54 m, mean age 29.1±3.7 yrs) data was used. The minimally preprocessed HCP resting-state data was used for this study. Additionally, the mean white matter, cerebrospinal fluid the motion regressors and their derivatives were regressed as mentioned by [1]. Moreover, different brain regions were identified using the Schaefer 400 atlas and for completeness 16 subcortical regions and 3 cerebellar regions were added using the HCP atlas ("Atlas_ROI2.nii.gz"). The voxel timeseries belonging to each of the 419 regions were averaged and a mean timeseries for every region was formed. Furthermore, a dynamic functional connectivity (dFC) matrix was formed using the sliding window technique. We used the window size of 576 sec out of the other explored window sizes (7.2, 36, 72, 144, 288 s) and stride was fixed at 7.2 s as it gave the best dynamic differential identifiability dIdiff score. To extract the SS information from dFC, Common Orthogonal Basis Extraction (COBE) algorithm was used. COBE attempts to find out common components (CC) in the dFC matrix and then the difference between the original dFC and the CC gives the SS components.
To quantify the SS components and show that it is better than the original dFC we use the metric proposed by [1] called (dIdiff). It assumes that dFC should be more similar between visits of the same subjects than between different subjects. It represents how accurately one can identify a new scan that belongs to a particular subject from a pool of subjects. The higher the dIdiff the more the intra-subject similarity and lesser the inter-subject similarity. dIdiff can be anywhere between 0 to 100, where a score of 100 signifies the best SS dFC that one could get. dIdiff gives a value for every time frame in the dFC, so we show only the max value here.
COBE algorithm must be first trained using the training data to get the basis that can obtain the CC (using which we obtain the SS components). We used single scans from 50 subjects to get the basis. Later during the test phase, the SS components were extracted from multiple scans of subjects that were not used during training phase (Note that the multiple scans were only used to quantify the SS using the dIdiff metric).
To quantify the SS components and show that it is better than the original dFC we use the metric proposed by [1] called (dIdiff). It assumes that dFC should be more similar between visits of the same subjects than between different subjects. It represents how accurately one can identify a new scan that belongs to a particular subject from a pool of subjects. The higher the dIdiff the more the intra-subject similarity and lesser the inter-subject similarity. dIdiff can be anywhere between 0 to 100, where a score of 100 signifies the best SS dFC that one could get. dIdiff gives a value for every time frame in the dFC, so we show only the max value here.
COBE algorithm must be first trained using the training data to get the basis that can obtain the CC (using which we obtain the SS components). We used single scans from 50 subjects to get the basis. Later during the test phase, the SS components were extracted from multiple scans of subjects that were not used during training phase (Note that the multiple scans were only used to quantify the SS using the dIdiff metric).
Results:
Single scans of 50 subjects were randomly selected for training the COBE algorithm and the other 50 subjects were used for test. This process was repeated 100 times to see any variability due to training subjects. We observed that the SS dFC computed by COBE gave a significantly high value of dIdiff as compared to that of original dFC (Paired sample t-test, t(99) = 295.78, p<0.001).
Conclusions:
Results show that COBE algorithm gives robust SS components when trained using only single subject dFC matrices. On average the max dIdiff score improves from 27.47 to 44.70 when COBE is used. This shows that the COBE algorithm has the potential to extract the SS components from dFC only by using single scan data which can be used for a variety of individual differentiability applications such as behavior prediction.
Modeling and Analysis Methods:
Connectivity (eg. functional, effective, structural)
fMRI Connectivity and Network Modeling 1
Task-Independent and Resting-State Analysis 2
Keywords:
FUNCTIONAL MRI
Machine Learning
Other - Dictionary Learning; Subject-Specific; dynamic Functional Connectivity; individual variation; Common orthogonal Basis Extraction;
1|2Indicates the priority used for review
Provide references using author date format
2) Finn, E. S., et al. (2017), ‘Can brain state be manipulated to emphasize individual differences in functional connectivity?’, NeuroImage, 160, 140-151.
3) Jain P., et al. (2023), ‘Enhancing the network specific individual characteristics in rs-fMRI functional connectivity by dictionary learning’. Human Brain Mapping, 8, 3410-3432.