Resting-state fMRI (rs-fMRI) has been widely used to test the functional network1changes in the investigation ofthe mechanism of Alzheimer's disease (AD) and mild cognitive impairment (MCI). Most of these studies analyzed the whole period of rs-fMRI data based on the long-term correlation between voxels however, the brain function changed very fast and the network states we obtained from 6-10min fMRI data integrate several shorter states. In recent years, the dynamic changes of brain functional network have become a trend topic both in neuroscience and neuroradiology fields but most of the work based on rs-fMRI recursively uses the correlation coefficient extracted from a small sliding window. In this study, we proposed a time point-based analysis method to observe dynamic pattern changes in normal control, AD and two MCI stages and investigate whereas the DMN deficit we observed in previous studiesis related to some changes of activity frequency of micro-state network condition.Four groups of subjects were enrolled in this study, including AD, late MCI, early MCI and age-matched NC. The group size of each group is 30. MRI exam included routine protocol for screening obvious lesions and a resting-state fMRI (rs-fMRI) session. The parameters of rs-fMRI are as follows: TR=2000 ms, TE=30 ms, flip angle=90°, 33 slices, slice thickness=3.5 mm, distance factor=20%, FOV=210 ×210 mm2, matrix= 64×64, measurements=200. All data were collected on a MAGNETOM Tim Trio 3T MR scanner (Siemens Healthcare, Erlangen, Germany).The pre-processing steps include: 1) removing the first 10 frames 2) slicetiming correction 3) head motion correction 4) spatial smoothing with Gaussian kernel, 6mm full width half maximum 5) removing the linear trend and regressing out 9 covariance parameters including global signal, white-matter signal, ventricle signal and 6 head motion parameters 6) projecting the volumetric data from each individual’s natural space to the standard surface space (fsaverage) using Freesurfer pipeline. For each subject, we extracted the pre-processed data from the DMNbased on a functional network template derived from 1000 normal subjects.2For each time point, a vector was constructed based on signal extracted from each vertex within the DMN. Then all vectors of the four subject groups were clustered into several groups using k-means algorithm. The number of groups was optimized according to the Davies-Bouldin index. In each group of vectors, the t-statistic value of each vertex was calculated across all time points to create a temporal state of DMN. The t-values with p>0.01 were set to 0.For each subject, the time points belonging to each cluster were counted. Then group analyses were used to test the differences in the appearance frequency for each separate cluster among groups.There are 4 clusters showingsignificant differences between groups (c3,c4,c5,C6). The topography of these clusters is shown in Fig.1 (left). The c1 and c2 are the clusterswith the highest appearancefrequency during resting-state fMRI exam forall subjects. From the topography of c1 and c2, we could see they are the standard activation and deactivation patterns of the whole DMN.The c3-c4, c5-c6 are also paired as activation and deactivation patterns that have almost the same appearance frequency in each individual subject.In the network state shownin c3-c4, the posterior cingulate (PCC) and lateral parietal cortex (LPC) werenot involved. And this states' appearance frequency showedsignificant decreases in all patient group.The other network state shownin c5-c6 demonstratedA micro-state of DMN that only involved the frontal lobe and a small part of LPC and mid temporal lobe areas of DMN.There are significantly higher percentagesof c5-c6 in patient groups than in normal control.Many previous studies havereported the differences of functional connectivity in PCC and LPC between NC and AD. If we average these different clusters across the whole experiments, our evidence agrees with previousseed-based or ICA-based analyses.The functional deficit reported in PCC and LPC may be because the synchronization betweenseveral sub-parts of DMN has been broken in AD.The dynamic patterns obtained by the time point-based analysis detected severalmicro-states of DMN. The appearance frequency of these micro-states will change during the cognitive impairment progress.