2.1 Data

All data for the individuals used in this analysis were collected from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). The ADNI was initiated in 2003 as a public-private partnership under the supervision of Michael W. Weiner, MD. The primary aim of ADNI has been to test whether positron emission tomography, magnetic resonance imaging, other clinical and biological markers, and neuropsychological test assessments can be combined for early Alzheimer’s disease prediction and for MCI progression. Demographic information, raw neuroimaging data, CSF measures, APOE genotypes, diagnostic information, and neuropsychological test scores are publicly available from the ADNI data repository (http://adni.loni.usc.edu). In this paper, a total of 217 subjects were used, including 53 AD patients, 103 MCI patients, and 61 healthy controls. Table 1 presents the demographics of all these subjects. All structural MRI scans for this paper were obtained from 1.5 T scanners.

The entries for age, gender, education, and MMSE denote mean and standard deviation for each group. MMSE, Mini-Mental State Exam; CDR, clinical dementia ratio.

2.2 Data Acquisition

We downloaded all the MRI images in Neuroimaging Informatics Technology Initiative (NifTi) format. Downloaded images were preprocessed for spatial distortion and B1 field inhomogeneity correction.

CSF data were collected in the morning after overnight fasting with a 20-G or 24-G spinal needle. Within one hour of acquisition, CSF biomarkers were frozen and transported to the ADNI core laboratory at the Medical Center of the University of Pennsylvania.

The ADNI biomarker core laboratory also provided genotype and gene expression data for each participant in this study, which were obtained from peripheral blood samples. The genetic feature was a single categorical variable for each participant, taking one of five possible values: (ε2, ε3), (ε2, ε4), (ε3, ε3), (ε3, ε4), or (ε4, ε4). In this study, we specifically analyzed APoE ε4 allele status (carrier vs. non-carrier).

2.3 FreeSurfer Analysis of MRI

We applied the recon-all FreeSurfer pipeline, version 6.0.0 (http://surfer.nmr.mgh.harvard.edu) to the MRI images for cortical reconstruction and volumetric segmentation [21]. This pipeline automatically generates reliable volume and thickness segmentation of white matter and gray matter, as well as subcortical volume. Subcortical volumetric segmentation and cortical reconstruction included removal of non-brain parts, Talairach-transformations, segmentation of subcortical gray matter and white matter regions, intensity standardization, and atlas registration. After these steps, a cortical surface mesh model was generated, and finally, the 34 cortical regions were obtained from cortical surface parcellation based on sulcal and gyral landmarks for both hemispheres corresponding to the Desikan-Killiany atlas [22].

2.4 Hippocampal Volume

Hippocampal segmentation was performed using the FreeSurfer [23] tool. Hippocampal volume is considered one of the major hallmarks in Alzheimer’s disease identification [24] due to the fact that detailed analysis of the hippocampus is considered a major step in the analysis. Thus, the hippocampus has been one of the frequently studied structures for diagnosing AD. However, this structure is not homogeneous, so it is usually subdivided into different subfields. Initial efforts to define the hippocampus subfields were mainly based on cell size, shape, and connectivity [25]. FreeSurfer separately segments the hippocampus subfields into right and left parts, which gives an estimate of the probability that every individual voxel associated with a certain arrangement is based on a priori insight regarding spatial relationships, which are obtained by using a training set. Differences in voxel intensity are located, and the subcortical region is parcellated, and then, affine registration in the Talairach space is performed. The detailed FreeSurfer processing stages are presented in [21], and the hippocampus subfield segmented regions are shown in Fig. 2.

Fig. 2. Hippocampal subfield segmented region from the FreeSurfer Free-view application.

2.5 Cortical Thickness

To calculate cortical thickness, T1‐weighted images were preprocessed using FreeSurfer [21]. To create high-contrast-to-noise-ratio images from brain-extracted images, intensity normalization was applied. The gray matter and white matter boundary was located by using images obtained after intensity normalization. After this, a triangular mesh around the white matter surface was constructed. Each brain hemisphere was then fragmented over 160,000 vertices of the triangular mesh. The mesh was outwardly deformed so the grey matter surface was created, and the boundary between the cerebral spinal fluid and grey matter surface followed the boundary. Cortical thickness was measured as the distance between the grey matter surface and the white matter surface for each vertex. A FreeSurfer common template was used to register the image using the cortical folding pattern of images. The neocortex was then parcellated into 68 neocortical brain regions (34 brain regions for each hemisphere) based on the Desikan-Killiany atlas as shown in Fig. 3. [22]. Mean thickness within that parcellation of all the vertices gives the thickness of each parcellation unit. Finally, 68 cortical thickness features were yielded per subject.

Fig. 3. Cortical thickness measurement using Free-Surfer (a) cortical region, (b) render surface.

2.6 Feature Selection

In most studies involving neuroimaging analysis, the number of predictor voxels obtained will outnumber the subjects. Thus, a dimensionality reduction technique is necessary in order to obtain the most relevant feature set, to discard noise and redundant features, and to escape numerical singularities and overfitting issues, thus enhancing classification efficiency. An efficient feature reduction algorithm is the essential section of a machine learning technique in cases of high-dimensional feature sets. We have shown the efficiency of the support vector machine-recursive feature elimination (SVM-RFE) algorithm in recognizing the early moments of AD [25]. Importantly, feature reduction was implemented only for the training data. Once identified, the same brain regions used during training were utilized to assess the predictive accuracy of the classifier on the test data. In this study, SVM-RFE was applied in order to obtain a ranked list of features that could best distinguish HC from AD and MCI. SVM-RFE is a multivariate wrapper technique-based feature selection algorithm, which precisely fits a model and eliminates the weakest feature until the defined informative number of features is reached. The SVM-RFE ranking criterion is closely similar to the SVM model. An SVM model is trained in every iteration of RFE. Then, the features with lower ranking criteria are eliminated, since they have the least effect on the classification process, while the remaining feature vectors are kept for the next iteration. This sequence is repeated until all the features have been removed. Then, in ranked order of elimination, the features are graded. A detailed description of the SVM-RFE model can be found in the literature. In this work, after the application of SVM-RFE, the highest ranked training features that maximize cross-validation accuracy were kept for training the classifiers.

2.7 Classification

The basic classification protocol for the proposed prediction framework is shown in Fig. 1. The machine learning framework consists of four major steps: feature extraction, feature selection, normalization, and classification. Brief descriptions of each of the classifiers used in this experiment are described here.

2.7.3 SVM

SVM is a supervised learning algorithm, and is one of the most widely known classification algorithms; its usage has been valuable in a large number of applications, including the classification and prediction of disease from structural MRIs of the human brain. Fig. 4 presents an SVM classifier that classifies data by constructing a hyperplane, which is defined as $w^{T} . x+b=0$, of a very high-dimensional feature vector, where $\textit{b}$ is the bias for the input vector, and $\textit{w}$ is the weight vector. A good identification is said to be the nearest training data sample belonging to any class—the greater the separation distance of the margin, the lower the generalization error of the classifier. Let us assume $\textit{N}$ training data samples $\left\{\left(x_{1}, y_{1}\right),\left(x_{2}, y_{2}\right), \ldots,\left(x_{N}, y_{N}\right)\right\}$ are given; $x_{i} \in \mathbb{R}^{d}$ is a set of feature vectors, and $y_{i} \epsilon\{-1,1\}$ is the class label. Then, the classification problem can be described by the minimization problem below:

(4)

$ \min _{w, b}\left(\frac{w^{T} w}{2}\right) $

(5)

$ \text { s.t. : } y_{i}\left(w^{T} . x_{i}\right)+b>1, i=1, \ldots n $

For a small dataset, an RBF kernel performs better than a linear kernel. A regularization constant, $\textit{C}$, and a set of kernel hyperparameters, $\textit{γ}$(gamma), need to be tuned in SVMs. Parameter optimization is obtained by using a cross-validation (CV) technique. This procedure is repeated 1500 times, each time randomly selecting a new set of 10 held-out subjects to obtain the optimum hyperparameter. In this method, the search scales for regularization constants and gamma values were set to C= [0.001, 0.01, 0.1, 1, 10, 100] and $\textit{γ}$= [0.001, 0.01, 0.1, 1, 10, 100], respectively. The maximum validation accuracy was obtained at $\textit{C}$=0.1 and $\textit{γ}$=0.01. The tuned parameters are used to predict the accuracy value on the test dataset.

Fig. 4. Support vector machine.

3.1 Classification Result

In this method, we utilized a binary classification technique to measure classification performance on the obtained cortical thickness, hippocampus volume, genetics, and CSF biomarkers.

Fig. 6 shows a comparison of AUC for AD vs. HC, HC vs. MCI, and AD vs. MCI. From the above ROC curve and Tables 2-4, we can see that the all-features combination with SVM classifiers outperformed the other classifiers, but for individual features, NB and KNN classifiers also performed better, compared to the SVM. For cortical thickness and CSF plus genetic measures in AD vs. HC classification, the KNN classifier had better AUC compared to the SVM and NB. Similarly, for hippocampus features in classifying AD vs. MCI, NB outperformed the other two classifiers in terms of AUC (91.65%).

3.2 Comparison with Other Methods

Recently, several research articles have analyzed neuroimaging and machine learning methods for Alzheimer’s disease diagnosis, with the focus on multimodality techniques. Table 5 shows a comparison of the classification performance by the proposed technique against recently published state-of-the-art methods that used multimodality datasets to diagnose Alzheimer’s disease. However, it is hard to compare the existing state-of-the-art methods, because the majority of the articles used various datasets and validation techniques, both of which influence classification performance. The study by Westman et al. [26] obtained accuracy of 91.8% with sensitivity at 88.5% and specificity at 94.6% for AD vs. healthy group classification of 369 subjects using the orthogonal partial least squares to latent structures (OPLS) method. Hinrichs et al. [27] obtained classification accuracy of 92.40%, with 86.70% sensitivity, 96.60% specificity, and 97.70% AUC when classifying AD vs. healthy patients for 230 subjects by using the MKL technique on multimodal ADNI datasets (i.e. sMRI, CSF, APoE ε4, and cognitive scores). In another study, Zhang and Shen et al. [28] using SVM classifiers obtained accuracy of 93.3% when classifying AD vs. healthy patients from among 186 subjects. Similarly, Liu et al., using multimodal cascaded CNN methods for sMRI and PET images [29], obtained accuracy of 93.26%, sensitivity at 92.55%, and specificity at 93.94%, with 95.68% AUC when classifying 397 subjects. Based on the existing literature, the proposed system (as results show in Table 2) was highly competitive in terms of classification performance, with accuracy of 96.93%, sensitivity at 99.89%, specificity at 95.40%, and AUC at 98.30% for AD vs. healthy patients.