Identification of MCI Individuals Using Structural and Functional Connectivity Networks

2.1. Subjects Characterization and Diagnosis

All subjects used in this study were recruited by the Duke-UNC Brain Imaging and Analysis Center (BIAC), Durham, North Carolina, USA. Informed consent was obtained from all participants, and the experimental protocols were approved by the institutional ethics board. This cohort involved of 27 participants, 10 individuals with MCI, and 17 socio-demographically matched normal controls. All the recruited subjects were diagnosed by expert consensus panels at the Joseph and Kathleen Bryan Alzheimer’s Disease Research Center (Bryan ADRC) and the Department of Psychiatry at Duke University Medical Center. Diagnosis was based upon available data from a battery of general neurological examination, neuropsychological assessment evaluation, collateral and subject symptom and functional capacity reports. Conformation of diagnosis for MCI if subjects met the following inclusion criteria: 1) age > 55 years and any race; 2) recent worsening of cognition, but still functioning independently; 3) Mini Mental State Examination (MMSE) score between 24 and 30; 4a) score ≤ −1.5 SD on at least two Bryan ADRC cognitive battery memory tests for single-domain amnestic MCI; or 4b) score ≤ −1.5 SD on at least one of the formal memory tests and score ≤ −1.5 SD on at least one other cognitive domain task (e.g., language, visuospatial-processing, or judgment/executive function) for multi-domain MCI; 5) 4 or lower for baseline Hachinski score; 6) does not meet the NINCDS-ADRDA or DSM-IV-TR criteria for dementia; 7) no psychological symptoms or history of depression; and 8) capacity to give informed consent and follow study procedures.

On the other hand, all healthy controls individuals met the following criteria: 1) age > 55 years and any race; 2) adequate visual and auditory acuity to properly complete neuropsychological testing; 3) no self-report of neurological or depressive illness; 4) shows no evidence of depression based on the Diagnostic Interview Schedule port based on the Diagnostic Interview Schedule portion of the Duke Depression Evaluation Schedule; 5) normal score on a non-focal neurological examination; 6) a score > −1 SD on any formal memory tests and a score > −1 SD on any formal executive function or other cognitive test; and 7) demonstrates a capacity to give informed consent and follow study procedures. In order for safety purposes and minimizing biases, subjects were excluded from the study if they have: 1) any of the traditional MRI contraindications, such as foreign metallic implants or pacemakers; 2) a past head injury or neurological disorder associated with MRI abnormalities, including dementia, brain tumors, epilepsy, Parkinson’s disease, demyelinating diseases, etc.; 3) any physical or intellectual disability affecting completion of assessments; 4) documentation of other Axis I psychiatric disorders; and 5) any prescription medication (or nonprescription drugs) with known neurological effects. Noteworthy that the diagnosis of all cases were made on clinical grounds without any reference to MRI. Demographic information of the participants is shown in Table 1.

2.2. Data Acquisition

Data acquisition was performed using a 3.0-Tesla GE Signa EXCITE scanner. Diffusion-weighted images of each participant were acquired axially parallel to the anterior and posterior commissures (AC-PC) line with twenty-five-direction diffusion-weighted whole-brain volumes using the following parameters: b = 0, 1000 s/mm², flip angle = 90°, TR/TE = 17000/78 ms, imaging matrix = 128 × 128, FOV = 256 × 256 mm², resulting in a voxel dimension of 2 × 2 × 2 mm³ reconstructed resolution. A total of 72 contiguous slices were acquired.

Resting-state functional images were acquired axially parallel to the horizontal plane connecting the AC-PC line with the following parameters: flip angle = 77°, TR/TE = 2000/32 ms, 34 slices, imaging matrix = 64 × 64, FOV = 256 × 256 mm², resulting in a voxel resolution of 4 × 4 × 4 mm³. A total of 34 slices were acquired using a SENSE inverse-spiral pulse sequence in the same plane as the low resolution T1-weighted images. All the subjects were told to keep their eyes open and stare at a fixation cross in the middle of the screen during scanning, which lasted for 5 minutes. As we know, neurons get excited to changing stimuli across time. But when the stimuli such as the little cross sign in this study was presented steadily without changing across the five minutes period, the neural excitation related to the stimuli can vanish quickly. Hence, this can ensure subjects not falling into sleep and avoid saccade-related activation which is unavoidable if eyes were closed.

2.3. Method

Progressive degenerative neurological diseases such as Alzheimer’s disease and similar dementias exhibit subtle, yet spatially and temporally diffuse pathology, where the brain is damaged in a large-scale, highly connected network, rather than in one single isolated region. In view of this, designing an informative description of interregional connections, which might be more sensitive in conveying the pathological information, is necessary for accurate diagnosis or prediction of neurological diseases.

It has been proven that information extracted from different modalities and different brain tissues might complement each other and provides more comprehensive characterization of brain abnormalities [9, 11, 12, 18, 36]. To integrate complementary information from different modalities and different brain tissue types, we proposed in this paper an approach that combine structural connectivity information from white matter (WM) regions and functional connectivity information from gray matter (GM) regions to distinguish individuals with MCI from normal controls. The key of our multimodality classification framework include:

• More informative description of the brain via multiple connectivity networks;

• Integration of DTI and rs-fMRI information at a whole-brain connectivity level;

• Multimodality data fusion using multiple-kernel SVM.

An overview of the proposed MCI classification framework is summarized graphically in Figure 1.

2.4. Summary of Methodology

The proposed multiple-kernel SVM based multimodality classification framework is summarized as follows.

1. In the DTI modality, whole-brain streamline tractography is performed using the 90-region AAL atlas to construct structural connectivity network based on the number of fibers passing through each pair of regions. Based on the common fibers connecting each region pair, on-fiber average FA, MD, and principal diffusivities are derived to construct another five connectivity networks.

2. For each network, the clustering coefficients for each ROI are determined to compute a total of 90 features. Feature vectors from all networks of each subject were concatenated to form a long feature vector to provide an informative description of WM connections finally resulting in a total of 540 features.

3. In the fMRI processing, the first 10 fMRI images are discarded to cater for magnetization equilibrium of each subject. Slice timing correction and intra-subject registration of the acquired rs-fMRI time series are performed on the remaining images using the SPM software package to correct the slice acquisition time delay and head motion.

4. T1-weighted image of each subject is segmented and its GM regions are then used to mask the acquired fMRI time series. The brain space is further parcellated into 90 regions using the AAL template.

5. The mean time series of each ROI, which is obtained by averaging all voxels within the GM-masked ROI, is band-pass filtered (0.025 ≤ f ≤ 0.100Hz) before it is further decomposed into five equally-divided, distinct frequency sub-bands. Five functional connectivity networks are constructed using pairwise Pearson correlation coefficients, one for each frequency sub-band.

6. For each network, the local clustering coefficients for all ROIs are determined as features for classification. All clustering coefficients are concatenated to form a long feature vector consisting of 450 features.

7. For each modality, features with significant discriminative power are selected using t-tests.

8. Individual kernel matrices are then constructed from the selected features of each modality before they are integrated to form a single mixed-kernel matrix.

9. The constructed mixed-kernel is employed to train SVM classifiers in a leave-one-out fashion using the LIBSVM solver.

3.1. Classification based on Multimodality Connectivity Maps

The proposed multiple-kernel SVM based multimodality classification approach was compared with the single modality approach and the direct data fusion method. In the single modality approach, only features selected from a single imaging modality (DTI or rs-fMRI) were applied for kernel matrix construction. In the direct data fusion method, all 990 features, including DTI and rs-fMRI features, were first concatenated into a long vector before feature selection and kernel matrix construction. In the multiple-kernel approach, the optimal weighting factor, β_m, was determined via grid search. Similarly, in all experiments parameters such as p-value for t-test feature selection method and SVM soft margin constraint parameter, C, were also determined via grid search.

Comparison was performed via leave-one-out cross-validation due to the limited number of available samples. Classification performance for identification of individuals with MCI using single and multimodality connectivity networks are summarized in Table 2. The proposed method yields a classification accuracy of 96.30%, which is an increment of at least 7.41% from that of the single modality approach and the direct data fusion method. It also outperformed all other methods in the rest of the computed statistical indices.

A cross-validation estimation of the generalization performance shows an area of 0.9529 under the ROC curve (AUC), indicating excellent diagnostic power. ROC curves for all compared methods are shown in Figure 4.

3.2. Effect of Weighting Factor, β_m

The weighting factor, β_m, determines the contribution of each modality in multimodality classification as formularized in equations (5), (6) and (7). A larger β_m value indicates larger contribution of DTI to the classification and vice versa. In this experiment, we seek to investigate how classification performance varies with respect to β_m. We fixed the constraint parameter, C to 3.0, and p-value to 0.01 for both modalities for feature selection. The classification accuracy and AUC value of the proposed method are shown graphically in Figure 5.

It is observed that higher classification accuracy is achieved when larger β_m value is used. Accuracy of more than 90% can be achieved with β_m equal or larger than 0.3. The highest accuracy values are achieved within the range of 0.40 ≤ β_m ≤ 0.65. Similar trends can be observed for AUC values. High and consistent AUC values are obtained over a relatively wide range of 0.35 ≤ β_m ≤ 0.70. This indicates that the proposed method is relatively robust to the weighting factor. The skewness of the accuracy and AUC graphs towards larger β_m values indicates dominance of structural information carried by the DTI modality in the classification.

3.3. Effect of Constraint Parameter, C

We investigate the influence of different values of the constraint parameter, C, on the classification performance. We fixed the weighting factor, β_m to 0.6 and varied the constraint parameter, C, from 1 to 10, in steps of 1. The p-value used for feature selection were maintained as 0.01 for both modalities. The classification accuracies and AUC values with respect to C are shown in Figure 6.

Classification performance is relatively robust to C. Variation of constraint parameter, C, does not significantly affect the performance of the proposed method. In addition, we also tested our method using several much larger C values, i.e., 50 and 100. The outcome again showed that the proposed method consistently maintains high classification accuracy, despite the high C values.

3.4. Effect of Predefined Threshold for Feature Selection

In the proposed framework, the statistical t-test is applied to select a subset of features with most discriminative power. Features with their p-value smaller than a predefined threshold are retained for construction of SVM classifiers. Thus, the number of features that will be selected is determined by the predefined threshold. In this subsection, an experiment is performed to explore the robustness of the proposed framework over the number of features selected. The performance is summarized in Figure 7.

A high classification performance is consistenly observed with the variation of p-value (variation in the number of selected features). Classification accuracy is more than 88.0% for most of p-values. For large p-values, classification accuracy deteriorates because more features, including some redundant and confounding features, are selected for constructing the classifier. The proposed framework shows a very consistent AUC value, always larger than 0.92, over the p-values, indicating excellent diagnostic power. The numbers of features selected for different p-values are provided in Table 3.

3.5. The Most Discriminant Regions

The most discriminant regions that are selected during classifier construction include the orbitofrontal cortex, temporal pole, anterior and posterior cingulate gyrus, frontal gyrus, amygdala, precuneus, thalamus, parahip-pocampal gyrus and insula, in line with results reported in previous studies. Regions which are selected by t-test feature selection method for mixed kernel construction are shown in Table 4. The top twelve selected regions used for multimodality classification are graphically shown in Figure 8.

3.6. Classification Performance Using Linear And Nonlinear Kernel Functions

In this subsection, we compared the classification performance of our proposed framework using the linear and nonlinear kernel-based SVMs. The best classification performance of the linear kernel as well as two other nonlinear kernels, i.e., radial basis function (RBF) and polynomial, are provided after choosing the optimal parameters via grid search. The polynomial kernel used in this study is a homogeneous kernel defined as

where T denotes matrix transpose, (x, y) denotes the training set of instance-label pairs and d denotes the polynomial degree. The best classification performance achieved for all compared methods are shown in Table 5.

It can be observed that the performance of polynomial kernel is the same as the linear kernel. In fact, we found that the best performance is found to be when d = 1. RBF kernel shows significant improvement in classification performance if compared to the linear and polynomial kernels. However, the proposed normalized linear kernel performs better than the RBF kernel. We adopted the normalized linear kernel in this study since choosing the appropriate value of σ for RBF kernel is very challenging in practice.

4.1. Significance of Results

Single modality of biomarkers has been widely employed for AD/MCI diagnosis and prediction, often achieving relatively high performance. Single-modality-based methods rely on simpler scanning protocols and hence require less image acquisition effort, making them more clinically feasible. However, information from different imaging biomarkers complements to each other and potentially improve prediction accuracy. Integration of biomarkers of different modalities with different data fusion methods for diseases diagnosis and prediction is still an open area of research. Our study demonstrates that multimodality connectivity network analysis can be employed to distinguish individuals with MCI from normal controls with high accuracy.

It is well know that misclassifying a healthy person to be a patient may be troublesome, however, misclassifying a patient to be a healthy person may cause severe consequences. If a patient is mis-diagnosed as a healthy person, necessary treatments to delay or cure the disease may not be provided on time during the critical treatment period. This will accelerate the progression of disease from mild to severe, a point where no effective treatment is available, eventually causing death of patient. Hence, it is tremendously crucial for a classifier to alleviate such circumstance by providing high sensitivity to the disease. Our results shown an improvement in sensitivity when both imaging modalities are combined through the multiple-kernel SVM algorithm. Statistically, our method performed better than any single modality method due to the improvement of sensitivity while maintaining high specificity. Furthermore, our method shows a relatively stable and robust performance with respect to the modality weighting factor and SVM constraint parameter.

The regions selected in the course of classification by our method are in agreement with previous studies and include the prefrontal cortex and or-bitofrontal cortex [14], temporal pole [13], anterior and posterior cingulate gyrus, precuneus and insula [8], amygdala [7], thalamus [30], parahippocampal gyrus [21, 4, 24], inferior temporal left, superior temporal gyrus right and precentral gyrus right [20], inferior temporal left, inferior frontal gyrus left, insula and superior temporal gyrus [17].

4.2. Methodological Issues/Limitations

The performance of our method may be affected by the unbalanced data between individuals with MCI and normal controls that was used for SVM model construction. This is typically the case in real world disease diagnosis, where patient samples are normally more difficult to collect, depending on the prevalence of the disease. A classifier will normally try to adapt itself for better prediction of the majority class; this will increase the overall classification accuracy. Although the sensitivity has been improved by using multimodality connectivity networks, the proposed framework at its current stage is not designed to handle this issue. Full investigation focusing on handling unbalanced data will be our future work.

Whole-brain connectivity network analysis, a relative new neuroimaging approach for disease diagnosis and prediction, is still facing some fundamental problems. One of the problems is the lack of standardized approaches for brain parcellation, resulting in network nodes that are often define either anatomically or randomly [35]. The scale of parcellation can affect network properties such as clustering coefficient, global and local efficiencies, path length, small-worldness, etc.

There are many other imaging modalities that have been used for AD/MCI classification and prediction, as mentioned in Section 1. Integration of imaging modalities other than DTI and fRMI will be explored in the future to further improve the classification performance based on the proposed framework. We also plan to apply our method to distinguish between MCI subjects who progressed to AD, and those who remained stable for a certain number of years.

Our current study is limited by the small sample and statistical power is hence a potential concern. Leave-one-out cross-validation, as employed in this study, provides an optimistic estimate of the classification accuracy since all except one of the subjects are used to train the classifier. Hence, almost all information available in the dataset is used in classifier model construction. For other classification approaches such as k-fold cross-validation, only N − k (N is the total number of subjects in the dataset) subjects are included during training process, and hence the performance obtained may be less optimistic to the small dataset used in this study. However, estimation of the classification accuracy using k-fold cross-validation might be more precise when compared to leave-one-out cross-validation provided that there are sufficient data to accurately train the classifier. Our dataset is quite diverse, and it includes both sexes and all ages between 55 to 84 for individuals with MCI and 55 to 88 for normal controls. However, the results obtained have to be verified in the future with larger datasets to reduce individual effects and to validate the effectiveness of the proposed technique.