2.1. Data acquisition and preprocessing
Participants in this study were selected from the Alzheimer’s disease neuroimaging initiative (ADNI) database (http://www.loni.ucla.edu/ADNI/). The images that were used for analyzing included: diffusion-weighted, T1 weighted, T2 weighted and, FLAIR scans of 72 participants. Subjects including 24 NC (Male (M)/Female (F) = 11/13, mean age = 75.32 ± 8.31 year, Mini-Mental State Examination )MMSE( = 29.04 ± 1.23), 24 AD patients (M/F = 16/8, mean age = 76.47 ± 8.23 year, MMSE = 20.17 ± 4.97) and, 24 MCI (M/F = 12/12, mean age = 76.04 ± 8.65 year, MMSE = 26.70 ± 2.07). Two MCI levels (early or late) are generally defined according to the Wechsler Memory Scale Logical Memory II(22). In this paper, we considered early and late groups that each of them contains 12 subjects to create MCI. NC subjects do not have a history of neurological or psychiatric disorders; subjects with AD who have signatures of the NINCDS-ADRDA (National Institute of Neurological and Communicative Diseases and Stroke/Alzheimer's Disease and Related Disorders Association) criteria which means AD's disorder. Also, MCI subjects have symptoms a subjective memory concern; However, they were without any significant disability in other cognitive domains (they follow everyday activities with any signs of dementia). All scans that used in this study, were acquired by a 3T MRI Scanner (GE Medical Systems scanner). We collected structural MRI from all participants: T1- SPGE: TR=6.9 ms, TE=2.8 ms, slice thickness=1.2 mm; T2: TR=650 ms, TE=20 ms, slice thickness=4 mm; and FLAIR: TR=11000ms, TE=147.9 ms, slice thickness=5 mm; also, we collected diffusion MRI from all participants: matrix size = 256 × 256 × 46; voxel size = 1.36 × 1.36 × 2.7 mm; the sequence contained 5 images acquired without diffusion weighting and with diffusion weighting along 41 non-collinear directions (b = 1000 s/mm²).
2.2. Structural MRI processing
We performed the processing of morphometrically and intensity level assessment by using the FreeSurfer image analysis pipeline (23) (version 6 on ubuntu 16.04) based on T1 and T2-FLAIR weighted images. The results of this process were divided into two parts that consisted of subcortical and cortical. In the first place, we performed the subcortical processing steps which included: volume, intensity pixel value, range of intensity, maximum, and minimum intensity pixel value in each region of atlas available in Freesurfer (Desikan-Killiany atlas). In the second place, the measurements related to cortical were done which included: thickness, surface area, gray matter volume and the curve in each region of Desikan-Killiany atlas from the right and left brain hemispheres. Part of the FreeSurfer pipeline conforms to the MRI scans to 1 × 1 × 1 mm3 resolution and corrects for bias field using the non-parametric non-uniform intensity normalization (N3) algorithm.
Briefly, this processing includes removal of non-brain tissue, automated Talairach transformation,surface deformation, tessellation of the grey matter-white matter boundary, topology correction, and segmentation (24, 25).
Finally, after doing all the processing in this section, 496 features were extracted from every subject and saved in CSV format.
2.3 Diffusion MRI processing
The diffusion MRI data were preprocessed using DSI-Studio software (developed by Fang-Cheng Yeh from the Advanced Biomedical MRI Lab, National Taiwan University Hospital, Taiwan, Supported by Fiber Tractography Lab, University of Pittsburgh, and made available at http://dsistudio. labsolver.org/Download/). For skull stripping and filtering the background region, we used the masks provided by DSI-Studio. Before DTI parameter measurement, head motion and eddy current effects were corrected using the DSI-Studio toolbox. We used two different reconstruction methods include model base (DTI) and free model (QSDR); with two different attitudes to process the diffusion images. Four general feature categories were extracted from the images; the first type of features was diffusion-based parameters in each region of the structural freesurferseg and Brodmann atlases, the second type of features was tract-based parameters in each region of the same atlases and the third type of features was structural connectome with two atlases that were mentioned above. Finally, the last type of feature was the graph network.
This process was performed for DTI model according to the following recommended parameters:
FA index was used to determine the fiber tracking threshold and Otsu’s method was used to set the anisotropy (FA) threshold to stop the fiber tracking, the number of seed points = 1,000,000, max angle = 60◦, step size = half of the spatial resolution, length constraint = 30–300 mm, and no spatial smoothing. Also, DSI studio software was implemented to calculate the connections and graphs of the "connectivity matrix and graph Network " tool with two "End" and "Pass" views.
In total, 92763 features were extracted from 72 participants. The features extracted of structural and diffusion processed MRI images in our work are given in the chart below. (Shown in Figure 1)
2.4. Feature selection and Classification
Feature selection means recognizing the most relevant features for pattern recognition and noise reduction that usually used in studies where the number of data is extremely high while few studied subjects are available The Fast Correlation-Based Filter (FCBF) method was better than the other models among all the feature selection models because less than quadratic time complexity (17). FCBF selected 66 features for the next step.
After the feature selection step, we performed four classification models including the DT, KNN, SVM and ANN with 5, 10, 15, 20, 30, 40, 50-fold cross-validation were used to evaluate the performance of the classifier. To evaluating the ANN model, training, validation and test phase was repeated for 30 times and finally the mean and standard deviation were reported. For this purpose, data set was divided into three parts: 60% for training, 20% for validation, and 20% for test. The MATLAB software R2012 (The Math Works, USA) was used to design and implement classification models.