This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TBME.2014.2323078, IEEE Transactions on Biomedical Engineering TBME-01499-2013.R1 1 Generation of the Probabilistic Template of Default Mode Network Derived from Resting-State fMRI Defeng Wang, Youyong Kong, Winnie CW Chu*, Cindy WC Tam, Linda CW Lam, Yilong Wang, Georg Northoff, Vincent CT Mok, Yongjun Wang, Lin Shi** Abstract - Default-mode network (DMN) has become a Youyong Kong is with the Department of Imaging and Interventional prominent network among all large-scale brain networks Radiology and the Research Center for Medical Image Computing, The Chinese University of Hong Kong, Shatin, NT, Hong Kong SAR, China which can be derived from the resting-state fMRI (rs-fMRI) (email: [email protected]). data. Statistical template labelling the common location of * WCW Chu is with the Department of Imaging and Interventional hubs in DMN is favourable in the identification of DMN Radiology and the Research Center for Medical Image Computing, The from tens of components resulted from the independent Chinese University of Hong Kong, Shatin, NT, Hong Kong SAR, China (email: [email protected]). component analysis (ICA). This paper proposed a novel Cindy WC Tam is with the Department of Psychiatry, North District iterative framework to generate a probabilistic DMN Hospital, Sheung Shui, NT, Hong Kong (email template from a coherent group of 40 healthy subjects. An [email protected]). initial template was visually selected from the independent Linda CW Lam is with the Department of Psychiatry, The Chinese University of Hong Kong, Shatin, NT, Hong Kong SAR, China components derived from group ICA analysis of the (email:[email protected]). concatenated rs-fMRI data of all subjects. An effective Yilong Wang is with the Department of Neurology, Beijing Tiantan similarity measure was designed to choose the best-fit Hospital, Capital Medical University, Beijing, PR China (email: component from all independent components of each [email protected]) Georg Northoff is with the Mind Brain Imaging and Neuroethics, subject computed given different component numbers. The Institute of Mental Health Research, University of Ottawa, Ottawa, ON, selected DMN components for all subjects were averaged to Canada (email [email protected]). generate an updated DMN template and then used to select Vincent CT Mok is with the Department of Medicine and the DMN for each subject in the next iteration. This process Therapeutics, and Lui Che Woo Institute of Innovation Medicine, The iterated until the convergence was reached, i.e., overlapping Chinese University of Hong Kong, Shatin, Hong Kong SAR, China (email: [email protected]). region between the DMN areas of current template and the Yongjun Wang is with the Department of Neurology, Beijing Tiantan one generated from the previous stage is more than 95%. By Hospital, Capital Medical University, Beijing, PR China (email: validating the constructed DMN template on the rs-fMRI [email protected]). data from another 40 subjects, the generated probabilistic **Lin Shi is with the Department of Medicine and Therapeutics, and Lui Che Woo Institute of Innovation Medicine, The Chinese University of DMN template and the proposed similarity matching Hong Kong, Shatin, Hong Kong SAR, China (email: mechanism were demonstrated to be effective in automatic [email protected]). selection of independent components from the ICA analysis results. I. INTRODUCTION Index Terms— Default mode network, resting-state fMRI, Over the last few decades, there has been growing interest template, brain network. in investigating the large-scale brain networks that exist in Manuscript received November 06, 2013; revised February 18, 2014; the human cognition, perception and emotion. The default- accepted April 30, 2014. This research was was supported by grants from mode network (DMN) is a prominent one among these the Research Grants Council of the Hong Kong Special Administrative discovered brain networks. The commonly accepted hubs of Region, China (Project No.: CUHK 475711, 411910, 411811), a grant DMN include the posterior cingulate cortex (PCC), the from the Science, Industry, Trade and Information Commission of Shenzhen Municipality (Project No. JC201005250030A), and a grant medial temporal lobes (MTL), the medial prefrontal cortex from the National Natural Science Foundation of China (Project No. (mPFC) and the angular gyrus (AG) [1, 2]. These related 81101111). brain regions are considered to be active during resting state while deactivated when specific purposeful cognitive tasks Defeng Wang is with the Department of Imaging and Interventional Radiology and the Research Center for Medical Image Computing, The are performed. Moreover, the DMN has been suggested to Chinese University of Hong Kong, Shatin, NT, Hong Kong SAR, China be related to various cognitive functions and be altered in a and CUHK Shenzhen Research Institute, Shenzhen, China (email: series of neuropsychological disorders [3, 4]. [email protected]). Copyright (c) 2013 IEEE. Personal use of this material is permitted. However, permission to use this material for any other purposes must be obtained from the IEEE by sending an email to [email protected]. 0018-9294 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TBME.2014.2323078, IEEE Transactions on Biomedical Engineering The identification of DMN has been extensively studied in produced by the ICA analysis on the concatenated 4D both healthy individuals and subjects of various dataset. However, the difference among the fMRI data of pathological conditions from resting state fMRI (rsfMRI) each subject may not capture the intrinsic functional [5]. Plenty of methods have been proposed to derive the connectivity [22]. The concatenated data can’t well reflect DMN from the rsfMRI datasets [6-8]. Among these the temporal relationship in the fMRI data of each subject, approaches, the method using independent component and thus the resulting template may not fully preserve the analysis (ICA) has been widely utilized to extract networks information in the original data. Moreover, only a few in previous studies [9, 10]. ICA technique enables the blind- subjects were utilized to generate the template [23]. source separation of mixed signals into independent spatial Therefore, it is of great importance to construct a and temporal components, which are considered to statistically representative template for DMN. represent different functionally connected networks [6, 11]. This paper presents a novel framework to generate a However, there remains the problem of how to accurately probabilistic DMN template from resting-state fMRI. select the component that best represents the DMN from the Inspired by the structural brain atlas generation, an iterative numerous components. Manual selection may be time approach was developed to create an unbiased DMN consuming as all components should be carefully visualized template. Resting-state fMRI data from 50 healthy subjects and analysed. One objective approach is to automatically were utilized to generate the template. An initial template select the best-fit component by matching with a given was visually selected from the independent components, DMN template [12, 13]. Therefore, a standardized DMN which were obtained by performing ICA analysis on the template representing the statistical location of the hubs is concatenated fMRI data from all the 50 subjects using FSL demanding for objective component selections. melodic software. For each subject, as the selection of Plenty of approaches have been so far proposed to construct component number could affect the ability to capture the the statistical structural brain template. In the early 1990s, resting-state functional connectivity, independent Evans, et al. [14] introduced the concept of a statistical MRI components were obtained on different component template for brain mapping. They constructed the MNI305 numbers. To find the DMN from the components in each template by linearly mapping 305 native MRI volumes to a subject, we developed a specific similarity measure. The manually-derived average MRI. In 1998, Holmes, et al. [15] component with the highest similarity was selected as the created a new atlas with much higher signal-to-noise ratio. DMN component for each subject. All the selected DMN One subject was scanned 27 times, and all the images were components were then averaged to generate a refined DMN linearly registered to compute an average which was finally template. The iteration process would stop once overlapping mapped to MNI305. However, linear registration can’t region between the DMN areas of current template and the capture the local deformation between different subjects. In one generated from the previous stage was more than 95%. 2000, Guimond, et al. [16] developed a method of building All selected components were finally masked with a given z a template atlas by nonlinearly registered each subject to an score. By averaging all the masks, a statistical