The study participants comprised 102 patients diagnosed with schizophrenia and 128 healthy controls that had been matched in age and gender. The chronic schizophrenic subjects were recruited from the Clinical Hospital of Chengdu Brain Science Institute (CBSI). The patients were diagnosed using the Structured Clinical Interview for the DSM-IV Axis I disorders—clinical version (SCID-I-CV), and all patients were being treated with medication (e.g., antipsychotics). The psychiatric symptom of schizophrenia was assessed using the Positive and Negative Symptom Scale (PANSS). The exclusion criteria for healthy controls were the following: a history of medical, neuropsychiatric illness, and major neurological or psychiatric illness in their first-degree relatives. This study was approved by the Ethics Committee of the Clinical Hospital of CBSI in accordance with the Helsinki Declaration. Written informed consent was obtained from each subject before the scanning.

Imaging was acquired on a 3T MRI scanner (GE Discovery MR750, GE Healthcare, Chicago, IL, USA). During scanning, foam padding and earplugs were used to reduce head motion and scanning noise, respectively. High-resolution T1-weighted images were acquired using a three-dimensional fast spoiled gradient echo sequence [repetition time (TR) = 6.008 ms, echo time (TE) = 1.98 ms, flip angle (FA) = 9°, matrix size = 256 × 256, field of view (FOV) = 256 × 256 mm², slice thickness = 1 mm, no gap, 152 slices]. Subsequently, resting-state functional MRI data were acquired using gradient-echo echo planar imaging sequences (TR = 2,000 ms, TE = 30 ms, FA = 90°, matrix size = 64 × 64, FOV = 240 × 240 mm², slice thickness/gap = 4 mm/0.4 mm, number of slices = 35), with an eight-channel phased-array head coil. All subjects underwent a 510-s resting-state scan to yield 255 volumes. The first 10 volumes (20 s) of data were discarded for magnetization equilibrium. During resting-state fMRI, all subjects were instructed to have their eyes closed and to move as little as possible without falling asleep.

Functional imaging data were pre-processed using the NIT toolbox (29) according to a standard pipeline similar with our previous study (30) and only briefly described here. Slice time correction and head motion correction were carried out. Then, the spatial normalization (3 ^* 3 ^* 3 mm) of the functional images was performed using 3D T1-based transformation. Any subjects who had a maximum translation larger than 3 mm or 3° were excluded. Framewise displacement (FD) was also evaluated as suggested by Power et al. (31). To minimize the effect in fMRI due to drifts, detrending analysis was performed. Then, sources of nuisance signals were removed from the normalized images through linear regression (six motion parameters and their first temporal derivative, white matter signal, cerebrospinal fluid signal). The global signal was not regressed out (32). Finally, fMRI data were passed through a bandpass 0.01–0.08 Hz.

Structural images were processed using the SPM12 toolbox. The whole-brain gray matter (GM) and total intracranial volume (TIV) were assessed, respectively. For each subject, the GM volume was normalized by dividing the TIV score. Then, the subjects' normalized GM value was entered as a global variable in statistical analysis to correct for the variability of global GM volume across subjects.

For each subject, we calculated sliding window-based whole-brain dynamic functional maps across time windows, including dynamic FCS and ALFF maps (Figure 1A). First, the whole run-time series were segmented into L windows, sliding the onset of each window by 10 s. Due to fact that the minimum window length should be no <1/f_min (28), s 100-s window length was selected (f_min = 0.01 Hz). Then, within the ith time window the FCS and ALFF maps were calculated. For each subject, we can get 40 FCS and ALFF maps, respectively. The calculation steps of FCS and ALFF were as shown below.

FCS analysis was performed as follows: (1) to exclude artifactual correlations from non-GM, the GM mask was generated by thresholding (cutoff = 0.2) the average of the GM probability map involving all subjects. (2) The time course within the GM mask was extracted, and the Pearson's correlations between every pair of voxels were calculated for each subject. (3) We then transformed individual correlation matrices to a z-score matrix using a Fisher r-to-z transformation. (4) For a given voxel (node), nodal FCS was computed as the sum of weights of its connections with other voxels. We conservatively restricted our analysis to positive correlations above the threshold of r = 0.2. Furthermore, to calculate the ALFF measure at each voxel within the GM mask, the time series of each voxel was transformed to the frequency domain through fast Fourier transform analysis. The power spectrum was then computed. The mean value of the square root of activity in the low-frequency band (0.01–0.08 Hz) was computed as the ALFF. Finally, individual voxel-wise FCS and ALFF maps were standardized by dividing the full-brain mean values and further spatially smoothed (FWHM = 6 mm), respectively.

For dynamic functional maps (FCS or ALFF), we tracked the FCS/ALFF-sDFN across time windows (Figure 1B) through group spatial ICA analysis (GIFT software: http://mialab.mrn.org/software/gift/). The low dimension was used to define the stable sDFN components. Twenty independent components (from the 1000 Functional Connectomes Project) (33) were initially extracted from all functional maps using the Infomax algorithm, which was repeated 20 times in ICASSO (http://research.ics.tkk.fi./ica/icasso). Nine sDFNs were firstly selected on the basis of the largest spatial correlation (SC) analysis with the corresponding template, including sensorimotor network (SMN), visual network (VN), auditory network (AN), DMN, salience network (SN), basal ganglia network (BGN), left frontoparietal network (LFPN), right frontoparietal network (RFPN), and dorsal attention network (DAN). The detailed steps were shown as the following: (1) One sample t-test was performed for each component. (2) The mask was calculated through family-wise error (FWE) correction. (3) For each template of RSN, Spearman correlation was performed between template matrix (1 by N matrix, N: the number of voxels within the whole brain) and 20-masked-component matrix (1 by N matrix, N: the number of voxels within the whole brain) respectively. (4) For each RSN template, the related sDFN was selected based on the largest SC score. Secondly, the selected nine components were visually inspected. Then, individual dynamic time courses and spatial maps of sDFN were reconstructed for each subject using the spatial-temporal regression approach. The dynamic time course of sDFN represents the variability of the functional network across the various time windows for each subject. The spatial map of sDFN represents the contribution of each voxel to the variability of the functional network across time windows.

A two-sample t-test was performed in each spatial map of sDFN between schizophrenic and healthy subjects [p < 0.05, cluster-level false discovery rate (FDR) corrected] with gender, GM value, age, and FD as covariates. In addition, the connectivity was computed between the dynamic time courses of two sDFNs by using Pearson correlations for each subject (Figure 1C). The Fisher r-to-z transformation of these correlations was performed. Two-sample t-tests were also performed for all potential connections among sDFNs with gender, GM value, age, and FD as covariates, with a statistical significance level of p < 0.05 corrected by FDR.

We examined the association between the PANSS scores and changed dFC within each sDFN and connectivity between sDFNs, respectively, with gender, GM value, age, and FD as covariates in schizophrenic subjects. FDR correction (p < 0.05) was used for multiple comparisons.

To assess the reproducibility of the spatial pattern of sDFN and the connectivity between sDFNs across a range of window sizes, sliding window-based FCS and ALFF calculations were performed in all subjects with a range of window sizes, respectively, including 100, 110, 120, 130, and 140 s. The score of SC was assessed between FCS/ALFF-sDFN and the corresponding network template. The connectivity between the time series of sDFNs was also calculated. Then, the ICC was computed for scores of SC and connectivity between sDFNs to quantify the reproducibility, respectively (34):

where σb2 represents between-subject variability and σw2 is the variability within subjects. An ICC value close to 0 indicates poor reproducibility, and a value close to 1 represents excellent reproducibility. Moreover, we compute statistical significance using the F-statistic (35):

where k is the number of repeated windows. The corresponding p-value of the F-statistic was then computed for degrees of freedom df 1 = 216 and df 2 = 868. Then, FDR correction (p < 0.05) was used for multiple comparisons for SC and connectivity between sDFNs, respectively.

Six schizophrenic subjects and four healthy controls with head motion scans exceeding 3 mm or 3° rotation were excluded. Additionally, three healthy controls were excluded because of deficiency of data by visual evaluation. Finally, 96 schizophrenic subjects and 121 healthy controls were included in this study. Schizophrenic subjects' clinical characteristics, including illness duration, medication dosage, and PANSS scores, are presented in Table 1. Both groups did not differ significantly in gender, age, and head motion (FD). The normalized GM of schizophrenic subjects was decreased compared to healthy controls (Table 1).

FCS/ALFF-sDFNs showed widespread functional architecture in both groups. As shown in Figure 2, the temporal variability of whole-brain voxels across the various time windows was spatially heterogeneous. We found that the FCS-sDFN showed nine specific patterns of network organization in the human brain, namely, SMN, VN, AN, DMN, SN, BGN, LFPN, RFPN, and DAN. ALFF-sDFN also showed eight specific patterns of network organization, namely, SMN, VN, AN, DMN, SN, LFPN, RFPN, and DAN. The BGN was not observed in ALFF-sDFN. Importantly, the spatial pattern of each sDFN was similar with the related RSN; the largest SC score is shown in Table 2.

The schizophrenic subjects showed decreased dynamic FCS/ALFF in sDFN compared to healthy subjects. Within FCS-sDFN, abnormal dynamic FCS was observed in BGN (thalamus), SN (anterior insula, anterior cingulate cortex, and supramarginal region), and SMN (post-central and pre-central gyri) in schizophrenic subjects. A similar reduced dynamic ALFF was also observed in ALFF-sDFN in patients with schizophrenia, including SN (anterior insula, and supramarginal region), SMN (post-central and pre-central gyri), and VN (calcarine and lingual regions) (Table 3 and Figure 3).

A significant connectivity between FCS/ALFF-sDFN, including positive and negative correlations, was observed in subjects. Depending on the connectivity pattern, these sDFNs could be divided into two clusters: positive connectivity within clusters and negative connectivity between clusters. One cluster comprised DMN, FPN, SN, and DAN; the other cluster comprised AN, SMN, VN, and BGN. In FCS-sDFN, a reduced negative connectivity was observed between LFPN and SMN and between BGN and RFPN in schizophrenic subjects (Figure 4). An altered connectivity was also found in ALFF-sDFNs in schizophrenia (Figure 4), such as decreased positive connectivity (between RFPN and SN) and a reduction of negative connectivity (between DAN and VN). Similar results were also observed in the other two window lengths (i.e., 120 and 140 s).

The relationship between clinical score and sDFN could further support the contribution of aberrant sDFN to the pathophysiology of schizophrenia. We observed a negative correlation between PANSS scores and dFC within the FCS-sDFN of schizophrenic subjects: PANSS-P score and insula within SN (r = −0.359, p < 0.005, Figure 5A1). We also found a negative correlation between the PANSS-total score and connectivity between ALFF-sDFNs (SN-RFPN, r = −0.369, p < 0.005, Figure 5A2) in schizophrenic subjects. No significant relationship was observed between PANSS scores and other altered dFCs. No significant relationship was observed between altered dFC and GM value, as well as FD score.

Both SC and connectivity between sDFNs have significant high reproducibility across a range of window sizes in subjects (Figure 6A: SC, Figure 6B: connectivity). The SC-ICC value of FCS-sDFN ranges from 0.45 to 0.76 (mean value: 0.66, standard deviation: 0.10). The connectivity–ICC value of FCS-sDFN ranges from 0.47 to 0.88 (mean value: 0.77, standard deviation: 0.08). The ALFF-sDFN ICC values of SC and connectivity between sDFN are higher than the ICC score of FCS-sDFN: SC-ICC: 0.75–0.87, mean ± standard deviation: 0.79 ± 0.04; connectivity–ICC: 0.93–0.96, mean ± standard deviation: 0.95 ± 0.007.

In the human brain, the functional connectors (e.g., fronto-parietal regions) may switch between different functional systems for higher cognitive flexibility during tasks (38, 39). Resting-state time-dependent functional networks may also reflect the spontaneous transitions among the potential functional coordinating configurations, which would provide a fast response to extrinsic stimuli (40). Empirical and modeling studies have been suggested that changes in static functional connectivity can be traced to changing patterns of communication dynamics (41). Thus, in the current study, we defined the sDFN through combining sliding window-based dFC and group ICA approaches to assess the spatial architecture of dFC across the various time windows in the human brain. We observed that the RSNs dynamically switched their functional networks across time, suggesting a time-dependent modular structure of the dFC in the human brain. The spatial maps of sDFN from ICA imply the contribution of each voxel to its sDFN. Specifically, these findings remained almost unchanged when using reproducibility measurements across lengths of window. The correction and interpretation of head motion should also be carefully handled in future studies because it may reflect individual variability in functional organization (42). These results suggest that the remarkable intra-subject sDFN observed here was not dominantly driven by the window length. Therefore, our novel findings highlighted that the sDFN during rest might form the intrinsic functional foundation for individual flexible task function or extrinsic stimulus.

Schizophrenic subjects showed decreased dFC within sensory and perceptual sDFN, including FCS (i.e., SMN) and ALFF (i.e., SMN and VN). Deficits of sensorimotor processing and multisensory functional connectivity integration, first investigated by Kraepelin and Bleuler, were well-documented as possible pathophysiological mechanisms in schizophrenic subjects (43, 44). Heightened dFC was then found within sensory and perceptual networks, such as VN (lingual gyrus and lateral occipital cortex), BGN, and SMN (pre-central and post-central gyri) in schizophrenia patients, which indicate that these perception regions are over local functional interacting across time (20). Furthermore, the increased time-varying connectivity of sensory and perceptual regions may result in spreading of disrupted internal and external sensory information to the distant high-order regions in schizophrenia (37, 45). In this study, we found decreased dFC in sensory networks (i.e., SMN, VN, and BGN) across the various time windows in schizophrenia patients. Combining our novel findings and previous studies, these decreased sensory sDFN may lead to reduced primary perceptual functional foundation and also contribute to the deficits of sensory high-order functional connectivity in schizophrenia.

SN was regarded as a whole unit to monitor internal and external salient information in the human brain (46). SN is one critical functional network that provides the basis of the representations of interoceptive awareness and external environment (46). Many symptoms (e.g., deficient facial affect processing and auditory affect processing) observed are involved with these aberrant static functions in schizophrenic subjects (47). Enhanced variability of functional connectivity within SN (e.g., insula and anterior cingulate cortex) in schizophrenia also supports the inappropriate function of sequential integration of interoceptive awareness with external perceptual stimulus in schizophrenic subjects (48, 49). This may eventually contribute to positive symptoms (e.g., hallucinations) (47). The deficient dFC of SN may reduce the functional foundation for regulating salient information and maintaining the integrated self in schizophrenic subjects.

In the present study, the deficient sDFN within primary perceptual regions and monitoring networks provided novel evidence to support altered dFC in processing external or internal sensory information in schizophrenia. These decreased dFCs may form deficient intrinsic function that may affect the intrinsic brain state, contributing to positive symptoms in schizophrenic subjects. Furthermore, recent studies hypothesized that abnormal sensory processing underlay cognitive impairment, in turn affected by high-order cognitive dysfunction (7, 8, 20). Unfortunately, the aberrant dFC between sensory and cognitive processes is still not regarded as the intervention target.

Neural processes depend on dynamic functional interactions between regions or networks in the human brain, which is thought to be instrumental for integrating and processing information in the course of behavior and cognition (41, 50, 51). These dFCs are coordinated through top-down projections from high-order functional networks, mostly located in FPN (52). Specifically, FPN could modulate sensory and other association networks to manage internal and external information (39, 53, 54). Combining with abnormal dFC within perceptual regions, altered static and dFCs of FPN revealed that dysfunctional modulation of FPN on internal states and salient stimuli may result in confusion about self-related and unrelated information in schizophrenic subjects (4, 9, 55, 56). In the present study, we did not find aberrant dFC within FPN in schizophrenic subjects, whereas decreased distant connectivity was observed in patients between FPN and distant networks, including SN, BGN, and SMN. Our findings may reveal that dysfunctional dFC between high-order and perceptual and monitoring systems may lead to less flexible function to modulate the internal and salient stimuli of schizophrenia.

In recent years, it has been hypothesized that the bottom-up sensory and perceptual dysfunctions may be a bottleneck in higher-level cognitive processing (7) and have been attributed as the causal role in clinical symptoms of schizophrenia (57, 58). Using the present advanced approach of sDFN, reduced dFC was consistently observed within primary sensory networks and SN, and between sensory and FPN in schizophrenic subjects. Critically, decreased dFCs of SN were tightly linked with severity of positive symptoms of schizophrenic subjects. The total PANSS score was also associated with decreased connectivity between SN and RFPN in schizophrenic subjects. Together, these findings verified the hypothesis that schizophrenia is more due to perceptual incoherence (2, 3, 7), which might be the causal role in clinical symptoms of schizophrenic subjects.