Analysis of Bipolar Disorder Using fMRI

In this work, a framework has been developed to analyse bipolar disorder using fMRI based on brain regional activity measurements. Eight brain regions have been considered in this study i.e. frontal lobe, temporal lobe, lentiform nucleus, insular, thalamus, caudate nucleus, parietal, and occipital regions. Initially, functional points are marked using independent component analysis and correlation coefficients established their connectivity. Based on the strength of the interregional connectivity, the activated points have been located the activated points using Hierarchical Modular Analysis and constructed network between the activated points of each brain regions. Extraction of Five feature descriptors i.e. centrality page rank and centrality degree, centrality closeness, assortativity and clustering coefficients have been done. Diagnosis operation is performed by heterogenous adaboost classifier. It was found that this work had given 94.2% accuracy which is comparatively better than earlier research works.


Introduction
Bipolar disorder is the sixth leading cause of disability in the world (WHO). Bipolar disorder, a serious mental health disorder is characterized [1] by periods of excitability or euphoria (mania or hypomania) alternating with periods of severe depression. Symptoms might bring about abnormal changes in mood and behaviour, which can cause serious distress and make life difficult. Hypomania and mania have the same symptoms. Mania is more severe than hypomania and results in more obvious issues with relationships, employment, school, and social activities. A psychotic break (psychosis) brought on by mania may also necessitate hospitalisation. Regarding records in advanced countries, about 2,40,000 in Australia [2], 7,25,000 withinside the UK, 3,90,000 in Canada, 8,10,000 in Iran, and nearly 10,00,000 in Germany have bipolar disorder (approx.). Bipolar disorder impacts as a minimum 1% of the populace, is related to accelerated mortality, and is a few of the pinnacle 10 maximum disabling ailments global. Currently, based on NIMH statistics report, the Prevalence Rate for bipolar disorder is approximately 1.1% of the population over the age of 18, or in other words, at any one time [3] as many as 51 million people worldwide suffer from bipolar disorder.
All regions of the brain can be impacted by bipolar disease, both physically and functionally. Bipolar disorder has a notable impact on the prefrontal cortex, grey matter, and hippocampus in the brain. The amount of grey matter in those with bipolar disorder was found to be considerably lower than in people without the disorder. The relationship between mood disorders and hippocampus volume was investigated by Wise et al. [4]. MRI images were utilised by Cao et al. [5] to examine the study participants' brains and calculate the volumes of various brain areas. They observed that bipolar disorder patients had the most pronounced size reductions in the hippocampus.
The motivations of this system are: (i) Time conservation in diagnosis and (ii) Exact diagnosing to reduce human errors while inspecting many images manually for diagnose. The overall objective is to diagnose BPD through medical images using image processing techniques, the specific objectives are (i) To improve the accuracy of detecting BPD over the recent identification methods. (ii) To develop a method to diagnose very faster than the existing methodology.
The main aim of this work is to analyse BPD by investigating the topological properties of functional brain networks within fMRI Signals of Healthy Control (normal) and BPD patients. The main contribution in this work is a hybrid technique to diagnose BPD in the earlier stage.
The highlights of this work are 1. Identification Brain components i.e. frontal lobe, temporal lobe, lentiform nucleus, insular, thalamus, caudate nucleus, parietal, and occipital regions 2. Local connecting connectome for each anatomical region. 3. Feature Extraction from functional connectome. 4. Heterogenous Adaboost classifier has been used for classification. 5. Developed a framework to identify BPD accurately in an early stage using Computer Aided Design.

Related Works
Clark et al. [6] had made studies of neurocognitive function in bipolar disorder in three domains: attention, executive function, and emotional processing. The prefrontal and anterior cingulate cortices, as well as subcortical limbic structures including the amygdala and the ventral striatum have been included in Functional imaging studies that implicated pathophysiology in distributed neural circuitry. Marchand et al. [7] had used functional connectivity analysis to identify circuits with aberrant connectivity. They found an increased functional connectivity among bipolar subjects in two cortical midline structures (CMS) circuits compared to the other health controls. Their results indicated that CMS circuit dysfunction persists in the euthymic state. Malhi et al. [8] reported that they had identified the brain regions associated with emotional processing in euthymic bipolar patients. It is noted that patients had less activation in the left ventral prefrontal cortex suggesting a potential trait deficit. Patients were slower to react than other health controls but did not differ in the rate of accuracy. Strakowski et al. [9] reported that bipolar patients had exhibited increased activation of limbic, paralimbic, and ventrolateral prefrontal areas, as well as visual associational cortices. Whereas Healthy subjects exhibited relatively increased activation in fusiform gyrus and medial prefrontal cortex. Chen et al. [10] showed that there is an abnormal frontal-limbic activation in BD. There was attenuated activation of the inferior frontal cortex or ventrolateral prefrontal cortex, which was consistent across emotional and cognitive tasks. Pavuluri et al. [11] reported that the pattern of reduced activation of ventrolateral prefrontal cortex and greater amygdala activation lead to negative stimuli. The pattern of functional alteration in affective and cognitive circuitry had contributed to the reduced capacity which affected regulation and behavioural self-control in paediatric bipolar disorder. Hilary et al. [12] had demonstrated the feasibility of investigating juvenile bipolar disorder using fMRI methods. Their findings suggested that the presence of front striatal abnormalities is common to adolescents and adults with bipolar disorder. Gruber et al. [13] suggested that differential processing strategies of bipolar patients had supported the theory of altered frontal systems in these patients during the performance of cognitive tasks.
Phillips et al. [14] reported that they made experimentation on the efficacy and tolerability of low and moderate dosages of extended-release quetiapine in adults with borderline personality disorder and stated that Participants treated with 150 mg/day of quetiapine had a significant reduction in the severity of borderline personality disorder symptoms compared to those who received placebo. Christoph et al. [15] found that there was a significant decreased frontal cortical volume and inferior frontal cortex in the Mania group, but no volume changes was noted in the No-Mania group. Blumberg et al. [16] reported that functional imaging studies of individuals at rest demonstrated regional abnormalities in primary mood disorders that were generally consistent in those who were predicted by the lesion studies.
Gong et al. [17] reported that they made multilevel kernel density analysis to identify brain network in which BD was linked to hyper-connectivity or hypo-connectivity and found Decreased GMV was found in the insula, inferior frontal gyrus, and anterior cingulate cortex.
Lawrence et al. [18] reported that they had compared with health controls and major depressive disorder patients. BPD patients demonstrated an increased subcortical and ventral prefrontal cortical response to both positive and negative emotional expressions. Pavuluri et al. [19] reported that a reduced activation in ventrolateral prefrontal cortex had reflected diminished top-down control that leads to the observed exaggerated activation in amygdala and paralimbic areas. Teng et al. [20] reported that there would be an alteration in resting-state connectivity associated with the striatal-thalamic circuit for bipolar patients. Rimol et al. [21] reported that the current surface-based methodology allowed a distinction between cortical thinning and reduction in cortical area and 1 3 revealed that the cortical thinning is the most important factor in volume reduction in bipolar disorder.
Selvaraj et al. [22] reported that bipolar disorder is consistently associated with reductions in right prefrontal and temporal lobe grey matter. Beyer et al. [23] stated that Microstructural changes in the white matter of the orbital frontal areas as reflected by increased apparent diffusion coefficient (ADC) values appear to be associated with bipolar disorder. Rubin et al. [24] stated that they had performed individual level classification of BPD or major depressive mood (MDD) in gray matter volume (GMV) based on Magnetic Resonance Imaging and a Support Vector Machine. The limitation in this study was this analysis included scans performed with two different head coils and scan sequences, which limited the interpretability of results in an independent cohort analysis. Bozzatello et al. [25] reported that their study was investigated to differentiate the brain activity patterns between BPD patients with identity diffusion and healthy controls using fMRI.
In this work, it is hypothesized that the analysis of bipolar disorder can be provided based on the regional brain activity measurements using fMRI of bipolar patients, Healthy Control and diagnose using adaboost classifier.

Methodology
The block diagram for the proposed work is given in Fig. 1. We have concentrated on eight brain regions in this research i.e. frontal lobe, temporal lobe, lentiform nucleus, insular, thalamus, caudate nucleus, parietal and occipital regions. Interconnected activity within 8 regions followed by feature extraction and classification operation is performed.

Data Set Description
The input fMRI images were taken from the OpenNEURO which is open science neuro informatics database storing datasets from human brain imaging studies research. The subjects taken for study include 49 patients with Bipolar Disorder and 130 in Healthy control where the subjects were in Resting State. In Bipolar, 22 Table 1 and the brain regions taken for analysis is given in Fig. 4.

Determination of Functional Connectivity
The functional points are selected in the brain region based on independent component analysis (ICA) on the input data using Fast ICA algorithm [26]. As per literature, using ICA, we have computed 90 [27] functional points. The algorithm type used in Fast ICA is  kurtosis which is the classical measure of non gaussianity. The kurtosis of y is classically defined by Finally, resultant matrix containing 90 independent components scaled to variance 1 of the input are computed. Then the functional connectivity is calculated between each pair of points 8 by using Pearson's correlation [20], resulting in a n|n correlation matrix for every participant which is represented as a correlation map in Figs. 5 and 6 for BPD and Healthy Control respectively.

Hierarchical Modular Analysis
The Functional Connectivity are further analysed by Hierarchical modular analysis (HMA) to determine the activated functional points based on the functional connectivity. In order to determine the modules with high to low interregional connectivity, HMA is done to cluster the n functional points into several modules based on the strength of the interregional connectivity in the activated functional points with reference to the work proposed by Teng et al. [20]. Figures 7 and 8 depict BPD and Healthy Control data set respectively and the activated functional points are depicted in different colours based on the clustered modules.
These clustered Functional Points are further used to determine the regional activity measurements.

Feature Extraction
The activated functional points which are obtained using HMA are taken for each region. In order to construct a network for each region to their obtained features for further classification, a graph is constructed as a network for the activated points in each region. The network plot is given in Fig. 9 in which for each region, the network is formed only for the activated functional points in its respective region.
Using the network of each region, the network metrics are attained to derive the network association of the individual brain region's functional points (Rutvi et al. [28]). The network metrics are calculated for the following parameters such as centrality page rank [29], centrality degree [29], clustering coefficient [30], assortativity [30], Centrality_closeness [30] for each region and then mean and standard deviation for page rank, degree, closeness and clustering coefficient are calculated.

Centrality Page rank
There are three main important factors that determine the centrality Page Rank of a node: the number of links it receives, the link propensity of the linkers, and the centrality of the linkers. The equation used is given in (2).
where D is a diagonal matrix, I is an N × N identity matrix, a is weight on the edges from vertex v.

Centrality Degree
It specifies the number of edges attached to the node and the formulae is given in (3).
where A is a matrix with vertices i, j where i ≠ j.

Clustering Coefficient
It shows the measure of the degree to which nodes in a graph tend to cluster together and the equation is given in (4).
(2) where d(v) is the degree of vertex v, N(v) is set of all nodes that are a distance 1 from a vertex v and A is the matrix.

Assortativity Coefficient
It is the Pearson correlation coefficient of degree between pairs of linked nodes. Positive values indicate a correlation between nodes of similar degree, while negative values indicate relationships between nodes of different degree. where e jk refers to the joint excess degree probability for nodes with excess degrees j and k. q k is a normalized distribution of a randomly selected node, given by q k = (k+1)P k ∑ j jP j and q is the standard deviation of the distribution q k .

Centrality Closeness
It signifies the measure of the average shortest distance from each vertex to other vertex.
where d (u, v) is the distance to all the other nodes in the network. With these network metrics and the number of activated points in each region, the analysis report is generated as a feature datafile and is given for the classification.

Classification
A meta-estimator known as an AdaBoost classifier [31,32] starts by fitting one classifier to the initial dataset before fitting further copies of the classifier to the same dataset. Which includes 3 classifiers GDA (Gaussian Discriminant Analysis), KNN (K-Nearest Neighbours), Logistic Regression and Naive Bayes for our work and implemented and automated using MATLAB (v.9, Mathworks, Inc.) software. Boosting is a "greedy" algorithm that in the end combines weak classifiers right into a strong classifier, and always accepts the extra classifier that strongly reduces the classification error at that precise iteration. The values of KNN parameters used in this work are Nearest neighbours: 30; In GDA and Naïve Bayes, the parameters are set to default values.

Discussion and Results
The proposed work was trained using Adaboost and performance of the algorithm was compared with an earlier work. The analysis was carried out with 49 bipolar and 130 healthy control fMRI Images. The performance was measured on 1.19 GHz Intel ® core™ i51035G1 CPU with 8 GB of RAM running Microsoft Windows 10 20H2. Table 2 represents the input data set used in this work.
We carried out a study and selected the 90 functionally defined points using Independent Component Analysis (ICA). For these points, functional connectivity is obtained by the calculation of 90 × 90 correlation map as in Fig. 10. The warm colours represent the positive correlations whereas the cool colours represent the negative correlations between functional points. The 90 points is represented in x axis(left-right) and in y axis (top-bottom) and the correlation value is utmost positive for the points with itself and thus they form maximum correlation for itself ie., in the main diagonal and the plot is mirrored in left and right of the main diagonal as the correlation coefficient from point a to b and b to a(say) is same.
The modular structure of brain functional connectivity is calculated for the correlation matrix which is signified in Fig. 11.
This modular pattern is obtained by reordering functional points in the correlation matrix according to maximizing the strength of connectivity close to the main diagonal of the correlation map.
For the activated functional points, regional activity measures are taken based on the network metrics that has been calculated based on the 8 parameters for the 8 regions in each dataset and the number of activated points in each region is also estimated and it is given in the Table 3.
Using the above 9 features obtained based on the region-based activity, the classification of Bipolar is achieved. The classification was completed with Adaboost classifier. In An AdaBoost classifier [21], we have used Logistic Regression as a base estimator. In this work, we have used n_estimators as 50 and learning rate as 1. The misclassified training samples get more weights, and the test error keeps decreasing even after 700 iterations. In Logistic Regression, Bias value as − 1.0059, Lambda value as 3.1674e−05 and Delta Gradient value as 1.4582 and the performance and the efficiency in determining the results was tested based on the cross-validation procedure of the input dataset   Table 4. Each feature is analysed with the performance measures (Mahdiyah et al. [33]). In identification of bipolar disorder, the efficiency and the performance are measured based on the accuracy, sensitivity, specificity, precision, recall, F_measure and Gmean. Results show that the Adaboost algorithm produces results that are optimal in performance compared to Bozzatello et al. [25]. During testing phase, we achieved 98.42% on sensitivity, 98.22% on specificity and 98.79% on recall and the accuracy of the proposed work is 94.2% and is more accurate when compared to work of Bozzatello et al. [25] which has accuracy around 80% and based on the  sensitivity and specificity ROC shown in Fig. 12 as well as AUC calculated based on ROC curve. We achieved AUC value as 97.49%.

Conclusion
The proposed work performed diagnosis based on the regional brain activity measurements using fMRI for bipolar or normal control. In this work, we have concentrated 8 brain regions which are usually affected by BPD as per literature. Our work has produced better accuracy of 94.23% when compared to the earlier work. This work can be extended for Parkinson and Alzheimer diseases too. Data Availability Data available on request from the authors.
Code Availability Software application.

Declarations
Conflict of interest the authors that they have no conflicts of interest to disclose.  Wiselin Jiji is a professor of computer science and engineering at Dr. Sivanthi Aditanar College of Engineering, Tiruchendur. She has published more than 78 scientific research papers. She is a recipient of 10 national and three state awards. Her long-term research focuses on Computer-Aided Detection (CAD) and Measurement (CAM) of lesions in medical images. CAD research aims at discovering the fundamental perception processes of human vision in the image-based diagnosis of lesions and developing mathematical/computational models that describes them. Her area of interests is computer-aided detection and diagnosis of abnormality using medical images and medical image analysis such as image enhancement, segmentation, feature extraction, object detection and pattern recognition. Catherine Praiseye Vijayan She doing Master of Engineering (ME) degree from Anna University, Chennai. Her research interest includes medical image processing and data mining.