Network analysis of resting-state fMRI (rsfMRI) has been widely utilized to investigate the functional architecture of the whole brain. The brain cortical and subcortical areas can be segmented or parceled into several functional and/or structural regions. The connectome analysis of human-brain structure and functional connectivity provides a unique opportunity to understand the organization of brain networks [1-4]. However, such analyses require an appropriate definition of functional or structural nodes to efficiently represent grey matter’s regions. Here we propose a robust parcellation method that first divides cortical and sub-cortical regions into sub-regions by clustering the rsfMRI data for each subject independently, and then merges those individual parcellations to obtain a global whole brain parcellation. To do so our method relies on majority voting (to merge parcellations of multiple subjects) and enforces spatial constraints within a hierarchical agglomerative clustering framework to define parcels that are spatially homogeneous.Forty right-handed subjects (23 males and 17 females, age=35 ± 10 years (average ± SD)) were enrolled in the study. T1-weighted images (TR/TE = 11/5 msec, FoV = 224x224 mm$$$^{2}$$$, voxel size= 0.54x0.54x0.9 mm$$$^{3}$$$) and rsfMRI data (TR/TE = 3000/30 msec, FoV = 131.8x131.8 mm$$$^{2}$$$, 50 slices, voxel = 2x2x3 mm$$$^{3}$$$ resolution, flip angle=90$$$^{0}$$$, 180 volumes) were recorded at 3T (Philips Achieva, AE Eindhoven The Netherlands) using a 32 channel head coil. Cardiac and respiratory processes were monitored using the scanner's built-in photoplethysmography and a pneumatic belt, respectively. rsfMRI data underwent a standard pre-processing, including removal of sources of physiological variance (heart beat and respiration), motion and slice timing correction and band-pass filtration (0.01–0.1 Hz). Finally, data were registered to a standard space (MNI 2x2x2 mm) and spatially smoothed (5x5x5 mm FWHM). The overall framework is visually described in Figure 1 (referring to parts A and B.) Two approaches on how to merge individual parcellation (A and B) are developed and compared with each other. Both approaches A and B start by hierarchically clustering brain function at each subject level, considering certain numbers of sub-parcels ($$$k$$$ = 100, 200, 500). Since our approach starts at the subject level, individual parcellations may not share similar labels across subjects which complicates the derivation of a group level connectome. Thus, we re-label each subject-level parcellation by measuring degree of overlap between a subject-level parcellation and a template. Two choices exist for the template. In approach A, we employ one of the subjects as a template. Approach B, relies on a group-level template for relabelling obtained by Independent Component Analysis (ICA) to reduce the dimension and with the Ward linkage agglomerative clustering technique [5]. To measure overlap we use the Dice metric [6]. Once relabeling is performed, a group-wise final parcellation is obtained by the majority vote across the study population. Reproducibility and cross-validation: We evaluate performance by measuring the degree of overlap of each clustering approach (A and B) with the AAL atlas. $$$R^{2}$$$ values can range from 0 to 1, with low $$$R^{2}$$$ value indicating lower reproducibility. $$$R^{2}$$$ is to measure how the regression fit to the leave-one-out model when each time remove one data point. The cross-validation was done by an approach that obtains group consistant parcels compare between AAL atlas.The results of the two majority voting rules are shown in the Figures 2A and 2B respectively. We investigate overlay percentage between AAL atlas and group consistent parcels in Figure 3. In Figure 3, the overlap scores show high reproducibility across 40 subjects especially in the high resolution case ($$$k$$$ = 500, DICE score = 86% ± 4%), demonstrating overall that to merge individual parcellations the template for relabeling does not play an important role. In general, our rsfMRI voting-based parcellation clustering can cover reach up to 90% AAL atlas especially when increasing the parcels $$$k$$$. We find out relied only on the individual subject merging into consistent parcellation without adding prior group-wise parcels can reach the similar level of the DICE score.We used rsfMRI data collected from 40 healthy subjects and we showed that our purposed algorithm is able to compute stable and reproducible parcels across the group of subjects at multi-resolution levels. Even though previous studies showed to ensure on average larger overlap between parcels and regions in standard atlas (AAL atlas [7]), the method proposed herein reduces inter-subject variability, especially when the number of parcels increases. Our high-resolution parcels seem to be functionally more consistent and reliable and can be a useful tool for future analysis that will aim to match the functional and the structural architectures of the brain.