Classification of Alzheimer’s Disease Using Whole Brain Hierarchical Network

 The main contributions of this paper can be summarized as follows:
            A hierarchical network is constructed to provide the complementary information for classification.
            New ROIS combinations and feature selection methods are also proposed.
            The MKBoost based classification method is used for AD and MCI diagnosis based on MRI data.

1.DATA

The cohort consisted of 200 patients with Alzheimer’s Disease (AD), 120 subjects that had mild cognitive impairment and converted to AD within 18 months (MCIc), 160 subjects with mild cognitive impairment that did not convert (MCInc) and 230 healthy controls (HC).

2. DATA PREPROCESSING

Firstly, non-parametric non-uniform bias correction (N3) algorithm  is used to correct intensity inhomogeneity. Secondly, the skull and cerebellum are removed by using BET . Thirdly, each MRI brain image is further segmented into three tissue types, namely gray matter (GM), white matter (WM) and CSF by using FAST . Finally, all brain images are affine aligned by FLIRT .

3.Hierarchical Network Construction

         In this article, we define the layers Lc,where c=1,2,3,4, to represent a brain of different number of ROIs as shown in Fig.3. The details of each layer are as follows:
       The layer L4 contains 90 original ROIs based on AAL atlas.
       The layer L3 contains 54 ROIs. A group of ROIs in L4 are merged into an ROI in L3 if their first two digits and the last digit are the same, for example, Cingulum Ant L (4001), Cingulum Mid L (4011) and Cingulum Post L (4021) are considered as an ROI in the layer L3.
        Similar to the process of the layer L3. A group of ROIs in L4 are merged into an ROI in L2 if their first and the last digit are the same, for example,Cingulum Ant L (4001), Cingulum Mid L (4011), Cingulum Post L (4021), Hippocampus L (4101), ParaHippocampal L (4111) and Amygdala L (4201)are considered as an ROI in the layer L2. The layer L2 contains 14 ROIs.

         The layer L1 contains only one ROI, that is the whole brain.

4.The Representation and Connectivity of the ROIs

         In this section, GLCM is used to extract 3D texture parameters of each ROI indirections.

         We then calculate the average texture parameters of each ROI in four directions.

         Afterwards, the connectivity between each pair of ROIs is evaluated by Pearsons correlation coefficients.

In this study, the F-score method is employed for feature ranking.

5.  Classification and Multiple Kernel Learning

In this paper, we use MKBoost algorithm proposed by Hao et al. [32] for classification tasks.

For each data set,we create a set of 13 base kernels, i.e.,

6. RESULTS AND DISCUSSION

  In this article, a 10-fold cross validation strategy is employed to evaluate the performance of the classifier.

Accuracy (ACC), sensitivity (SEN), and specificity (SPE) are used to quantify the performance of the classifier based on the results of 10-fold cross validation.

where TP, FP, TN, and FN are the numbers of true positives,false positives, true negatives, and false negatives, respectively.

In addition, we also report the area under receiver operating characteristic (ROC) curve (AUC). The AUC value is an important index to measure the overall-performance of classification methods, and the higher the AUC value, the better the performance of the classification method.