Resting-state functional magnetic resonance imaging: features of statistical processing of ROI-analysis data

Full Text

BACKGROUND

Resting-state functional magnetic resonance imaging (fMRI) is used to study the pathophysiology of diseases associated with impaired functional brain activity [1–3] by showing the degree of blood oxygen saturation in a local area of the brain (BOLD signal). According to recent research, pathological connectivity between resting neural networks may underlie various neurological and psychiatric diseases [4, 5]. Resting neural networks refer to spatially independent brain structures that are capable of coordinated activations without specific tasks or stimuli and involved in performing various cognitive functions; detection, processing, and integration of internal and external stimuli; and mental processes [6, 7]. Modern neurobiological studies described eight main resting neural networks [8–10]:

1. Passive mode network (DefaultMode): medial prefrontal cortex, posterior cingulate cortex/preclinical, and inferior parietal lobe on both sides
2. Sensorimotor network (SensoriMotor): superior (supplementary motor cortex of medial sections of frontal lobes) and bilateral lateral (motor and sensory cortex on both sides)
3. Visual network: medial, occipital, and lateral subnetworks
4. Significance network (Salience): anterior cingulate cortex, insula, rostral prefrontal cortex, and supramarginal gyrus
5. Attention network (DorsalAttention): frontal visual field and intraparietal sulcus
6. Executive control network (FrontoParietal): dorsolateral prefrontal and posterior parietal cortex on both sides
7. Language network: inferior frontal gyrus and posterior superior temporal gyrus on both sides
8. Cerebellar network: anterior and posterior cerebellar lobes

Several studies used the neural network construction method for region-of-interest-based analysis (ROI-analysis) to examine intra- and inter-network connectivity. Inconsistent findings in assessing brain connectivity using ROI-analysis can be attributed to methodological differences in statistical processing of fMRI data. To obtain valid scientific data, reproducibility of results should be established. The primary distinctions are evident in selecting the statistical analysis method for neural networks.

The most commonly used methods for statistical data processing in ROI-analysis are parametric statistics based on functional network connectivity, nonparametric statistics based on randomization/rearrangement analysis (i. e., spatial pairwise clustering), and nonparametric statistics based on threshold-free cluster enhancement [11].

Currently, there are no scientific studies that apply several statistical processing methods to one dataset. Therefore, a study that includes a comparative evaluation of different statistical methods for ROI-analysis in the processing of resting-state fMRI data should be conducted.

The present study aimed to assess the functional connectivity of the primary resting brain neural networks in ROI-analysis using various statistical methods.

MATERIALS AND METHODS

Fifteen resting-state fMRI studies of the brain from the Department of X-ray and Radiology (with a course of ultrasonic diagnostics) of the Military Medical Academy were retrospectively analyzed. MR scans of patients without neurologic and psychiatric pathology were selected based on analysis results. Functional MR scanning was performed using a Phillips Ingenia 1.5 Tesla scanner in the MRI department. Resting-state fMRI images were acquired using EPI-BOLD, with a repetition time (TR) of 3000 ms, an echo time (TE) of 50 ms, a field of view of 250 mm, a flip angle (FA) of 90°, a matrix of 128 × 128, a slice thickness of 4 mm, a slice gap of 0.6 mm, and 280 volumes. High-resolution T1-weighted structural images were obtained using a 3D-TFE sequence with a TR of 7.5 ms, a TE of 3.5 ms, an FA of 8°, a matrix of 256 × 256, and a slice thickness of 1.2 mm for volume 1.

The data were analyzed using neuroimaging software, including MATLAB, CONN21a, and SPM 12 [12]. Preprocessing involved functional alignment and unfolding, slice synchronization correction, direct functional segmentation, and normalization in the Montreal Neurological Institute space coordinate system. Further, functional spatial smoothing was performed using an 8-mm-wide Gaussian kernel. Outliers, which are BOLD signal artifacts, were identified using the ART toolkit if they differed by more than 3 standard deviations from the mean image intensity. To exclude false noise sources (e.g., physiological), an anatomical component-based noise reduction strategy (aCompCor) was employed [13]. Finally, a band-pass filter with a frequency window of 0.01–0.1 Hz was applied.

Following all preprocessing procedures, ROI-analysis was conducted to establish the functional connectivity between neural networks in the brain. This was evaluated by calculating correlations (using the Pearson correlation coefficient) between the time series of BOLD signal changes throughout the brain.

The following were used for statistical processing of ROI-analysis data:

1. The functional network connectivity (FNC) method is based on multivariate parametric statistics [14]. The process begins by identifying the networks of interest. Then, FNC analyzes all links between pairs of ROIs within and between networks by performing a multivariate parametric analysis of the general linear model of all connections included in each of these sets (clusters) of links. To evaluate individual clusters, an adjusted p-value corrected for the expected false discovery rate (FDR; Benjamini–Hochberg coefficient) at the cluster level (p < 0.05) was used to select significant interconnect sets. Additionally, an unadjusted p-value for the height (connection level) threshold (p < 0.05) was utilized to characterize the structure of individual connections within each significant set.
2. The spatial pairwise clustering (SPC) method [15] is a nonparametric statistic based on randomization and rearrangement. The process begins with a matrix estimated using a general linear model. The networks in this matrix are sorted either manually by the user (e.g., from an atlas) or automatically using a hierarchical clustering procedure [16]. Then, a statistical parametric map is determined using a height threshold (p < 0.001). The suprathreshold regions that result define nonoverlapping clusters. Each cluster is characterized by its mass, which is the sum of the square of the F- or T-statistics for all compounds within the cluster. These masses are compared to the distribution of expected cluster mass values under the null hypothesis. The null hypothesis is numerically estimated using multiple iterations of randomization/rearrangement of the original data. To evaluate individual clusters, an unadjusted p-value at the cluster level (p < 0.01) and a p-value adjusted for expected FDR at the cluster level (p < 0.05) were used to select only the significant clusters.
3. The threshold-free cluster enhancement (TFCE) method [17] is a nonparametric statistic. Similar to SPC analysis, TFCE starts with the entire matrix estimated using a general linear model, with the networks sorted either manually or automatically. Instead of defining a parametric statistical map using a height threshold, the analysis proceeds by computing the corresponding TFCE score map, combining the strength of the statistical effect for each connection with the size of all clusters. The null hypothesis’s expected distribution of TFCE values is then numerically estimated using ≥1000 iterations of randomization or rearrangement of the original data. A cluster level adjusted family-wise error p-value (p < 0.05) is applied to select significant clusters.

RESULTS AND DISCUSSION

Based on the results of various statistical processing methods, connectivity matrices that allow for the evaluation of inter- and intra-network connections were obtained (Figs. 1 and 2). The FNC method reveals that the visual and SensoriMotor networks exhibit positive connections with each other and with the DorsalAttention network and negative connections with DefaultMode and FrontoParietal networks (Fig. 1). The DorsalAttention network displays positive connections with the Salience, visual, and SensoriMotor networks and negative connections with the DefaultMode, FrontoParietal, language, and cerebellar networks. The cerebellar network has only one connection. The Salience network interacts positively with the DorsalAttention, FrontoParietal, and language networks. The language network is positively functionally related to the Salience and FrontoParietal networks and negatively related to the DorsalAttention network. Moreover, the language network has connections with both direct and inverse correlation with the DefaultMode network. The DefaultMode and FrontoParietal networks are interconnected with the DorsalAttention, visual, SensoriMotor, and language networks. Additionally, the FrontoParietal network is linked to the Salience network.

 

Fig. 1. Connectivity matrix using the method of multivariate parametric statistics based on functional network connectivity (FCN)

Рис. 1. Матрица коннективности при использовании метода многомерной параметрической статистики на основе функциональной сетевой коннективности (FCN)

 

Fig. 2. Connectivity matrix using non-parametric statistics methods based on: а — randomization/permutation spatial pairwise clustering (SPC); б — threshold free cluster enhancement (TFCE)

Рис. 2. Матрица коннективности при использовании методов непараметрической статистики на основе: а — анализа рандомизации/перестановки — пространственная парная кластеризация (SPC); б — улучшения беспорогового кластера (ТFСЕ)

 

The nonparametric statistical methods applied revealed that the visual, SensoriMotor, and DorsalAttention networks do not have any connections with the DefaultMode and FrontoParietal networks (Fig. 2). Furthermore, the DorsalAttention network does not have any connectivity with the language and cerebellar networks. Finally, the DefaultMode network is only connected with the FrontoParietal network. The networks described above have identical connectivity at SPС and TFCE. However, in the other networks, the results differed (Table 1).

 

Table. Connectivity of brain networks using nonparametric statistics methods Таблица. Коннективность нейросетей головного мозга при использовании методов непараметрической статистики
Resting neural networks	SPС	TFCE
Salience	DorsalAttention, Language, FrontoParietal	DorsalAttention, Language
Language	Salience, FrontoParietal	Salience
Cerebellar	FrontoParietal	–
FrontoParietal	Cerebellar, SensoriMotor, DorsalAttention, Language, DefaultMode	DefaultMode

 

The FNC method recorded 280 connections between brain network structures, whereas SPC and TFCE showed 186 and 182, respectively. Negative correlations were only found when parametric statistics were used (Fig. 3) and are significant as they indicate the suppression of one network while activating another. The negative correlation of networks is supported by the increasing popularity of the triple network model theory [5].

 

Fig. 3. Spatial image of the functional connections of the resting state networks of the brain using processing statistics methods: а, г — functional network connectivity (FCN); б, д — randomization/permutation spatial pairwise clustering (SPC); в, е — threshold free cluster enhancement (TFCE)

Рис. 3. Пространственное изображение функциональных связей нейросетей покоя головного мозга при использовании методов статистики обработки: а, г — функциональная сетевая коннективность (FCN); б, д — анализ рандомизации/перестановки — пространственная парная кластеризация (SPС); в, е — улучшение беспорогового кластера (ТFСЕ)

 

Several studies of brain neural networks did not specify the statistical processing method used. Divergent results of resting-state fMRI data may be related to different methodological approaches. The present study demonstrated that the choice of a statistical method for a single dataset significantly affects study results.

Although parametric statistics reveal a greater number of relationships, whether they have an advantage over nonparametric methods is unclear. Further research is warranted to determine the information these methods provide in intergroup processing, particularly when comparing fMRI data between patients with any pathology and controls.

CONCLUSIONS

Resting-state fMRI has become increasingly popular for studying the functional activity and connectivity of the brain. However, there are currently no standardized algorithms for constructing neural networks. Various methods are available for analyzing the functional integration and segregation of the brain, each of which provides unique information. Remarkably, each method has its distinctions in statistical data processing.

The use of multivariate parametric statistics in the FNC method enables the identification of a greater number of functional relationships compared to nonparametric methods. This method is particularly useful in identifying links with negative correlation. Our study determined that SPC based on randomization/rearrangement analysis was more informative than the threshold-free enhancement method among nonparametric methods.

ADDITIONAL INFORMATION

Funding source. Funding for this work has not been carried out.

Authors’ contributions. All authors contributed substantially to the study and article and read and approved the final version before publication.

Conflict of interest. The authors declare that there are no obvious and potential conflicts of interest related to the publication of this article.

Ethical review. The study was performed as part of a dissertation research and approved by the local ethical committee.

About the authors

Shamil’ K. Abdulaev

Military Medical Academy

Author for correspondence.
Email: izvestiavmeda@mail.ru
ORCID iD: 0000-0002-5126-4212
Russian Federation, Saint Petersburg

Dmitriy A. Tarumov

Military Medical Academy

Email: izvestiavmeda@mail.ru
ORCID iD: 0000-0002-9874-5523

MD, Dr. Sci. (Medicine), Associate Professor

Russian Federation, Saint Petersburg

Kirill V. Markin

Military Medical Academy

Email: izvestiavmeda@mail.ru
ORCID iD: 0000-0002-6242-1279
Russian Federation, Saint Petersburg

Aleksandra А. Ustyuzhina

Military Medical Academy

Email: izvestiavmeda@mail.ru
ORCID iD: 0009-0003-7282-0163
Russian Federation, Saint Petersburg

References

Supplementary files