Alzheimer’s Disease Classification Using Wavelet-Based Image Features

Alzheimer’s disease (AD) is a neurological syndrome affecting a large population of early-aged people worldwide. AD results in a continual decline in cognitive and communication skills which makes an individual unable to work without assistance [1]. More than 5 million people in the United States alone, were suffering from AD in 2018, and it is expected that this figure will be 15 million by 2050. The AD is not limited to the US people only, but it is spreading on a large scale across the world. The main reason behind the progression of AD is the accumulation of amyloid and tau proteins in brain regions. This accumulation causes synaptic loss in the brain and results in structural changes in the brain such as ventricular enlargement, hippocampal volume and size variations, cerebral cortex surface contraction, and grey matter density fluctuations [2, 3]. Magnetic Resonance Imaging (MRI) is one of the best neuroimaging tests that captures these structural changes with good tissue contrast and high resolution [4]. This paper presents a new feature extraction technique that extracts directional local descriptors which depict the structural and grey matter density variations at different stages and are helpful in early AD detection.

In literature, most of the studies include AD detection methods based on Voxel-based morphometry (VBM) [5], Surface-Based Morphometry (SBM) [6], Region of Interest (ROI) [7], Transform-based methods [8] and Texture-based methods [9]. VBM technique provides grey matter density details and compares the anatomy of different brains with a template. This technique suffers from the problem of imperfect image registration and misalignment issues of template and image [10]. The SBM provides cortical features but requires skills in the Freesurfer tool. This technique faces challenges in perfect spatial registration and surface reconstruction which affect the reliability of morphometric measurements [11]. ROI-based methods require segmentation to extract the affected part of the brain, and this needs expertise with prior knowledge. Transform-based methods generally suffer from high dimensionality and are less directionally sensitive to 3-D images [11]. Texture-based methods face challenges in clinics due to the lack of standardized approaches for acquiring MR images, performing intensity discretization on MR images, and selecting MRTA software [11]. Liu et al. [12] used the VBM technique to obtain voxel-wise grey matter density maps from different local patches for AD classification. The authors used a single atlas for obtaining patches which is mostly inclined for a particular class. In ROI-based studies [13, 14] hippocampal visual features [13] and tissue-segmented features [14] have been used for AD classification. These techniques provide spatial domain features which are not directional. The recent studies [15-17] are transform domain studies. These studies [15-17] used shearlet transform [15] and contourlet transform [16, 17] to obtain features from different ROIs and density maps for AD classification. These studies [15-17] included the limitations of VBM and ROI-based methods. However, these studies [15-17] provided frequency domain features which show multiscale and directional information. Most of the texture-based studies used local descriptors as features for AD classification. Bhasin et al. [18] applied 3-DWT to obtain Local Binary Pattern-20 (LBP-20) features for AD classification. This technique used DWT which has low directional selectivity in comparison to complex wavelet. Francis and Pandian [19] proposed a feature extraction technique which enhances the effectiveness of a fast Hessian detector by combining it with the local binary pattern. Sarwinda and Bustamam [20] combined 2D and 3D advanced LBP for multi-class classification of Alzheimer’s disease. This method suffered from high dimensionality and large computational time. Oppedal et al. [21] combined the LBP texture features with the contrast measures extracted from MR scans and obtained 98% accuracy. Koh et al. [22] applied bidirectional empirical mode decomposition on MR images and obtained four IMFS. The authors computed LBP histograms from these IMFS and further used these histograms for AD classification. Kaplan et al. [23] obtained Histogram-oriented gradients (HOG), local binary pattern (LBP) and local phase quantization (LPQ) from brain images. The authors merged all features and selected optimum features by using Neighbourhood Component Analysis for AD classification. All the discussed techniques [18-23] are texture-based and extract features based on textures only.

This paper instigates an approach that combines the characteristics of transform-domain and texture-domain techniques for extracting features. These features have the characteristics of both spatial domain and frequency domain and thus are more informative. The contribution of the paper can be listed as follows. (1) The proposed method utilizes DD-DTCWT for extracting features in sixteen directions. Thus, the proposed technique provides directional features. (2) This paper proposed a method which utilizes the local descriptors having the properties of both circular and elliptical LBP’s. Circular LBP provides isotropic information and Elliptical LBP provides anisotropic information. In tradition, the histograms of CLBP and ELBP need to be concatenated to capture both isotropic and anisotropic details in the image. This increases the feature vector size. In the proposed technique, CELD is used which provides circular and elliptical LBP properties with just half of the feature vector size [24]. (3) The proposed method can be applied for the detection of different stages of AD as it captures both grey matter density fluctuation and multi-structural variations in the brain with the advancement of the disease. Imaginary coefficients capture structural information and detail sub-bands provide information regarding grey matter density fluctuation. (4) This paper uses the shifted version of WSCELD. This contributes to providing several adaptable micro and macro patterns. (5) This paper compares the performance of different versions of WSCELD like Median WSCELD, Mean WSCELD, Energy WSCELD and Variance WSCELD using different wavelets like DD-DWT [25], DTCWT [26] and DD-DTCWT [27].

The remaining paper is divided into three sections as follows. The second section includes the proposed approach with background material. The third section includes the details of the database, performance metrics and performance discussion of the proposed work and existing methods. The fourth section throws light on the conclusion with future directions.

3.1 Dataset

This study used Open Access Series of Imaging Studies (OASIS) dataset for executing the proposed algorithm [33]. OASIS consists of 3-D MR scans of three categories AD, M_iCI and N_oC having Clinical Dementia Rating (CDR) value of 1,0.5 and 0 respectively. Total 84 3-D MR scans with 28 scans belonging to each category have been obtained for implementation work. Subsequently, 336 MR slices with 112 slices per category are taken out from all 3-D MR scans. These 336 MR slices are the centre slices. The sample images of the centre slices are shown in Figure 7.

7.png

Figure 7. Sample from OASIS dataset (a) Alzheimer’s Disease (AD) (b) Mild Cognitive Impairment (M_iCI) (c) Normal Control (N_oC)

3.2 Performance test criteria

The performance of the proposed work is evaluated using three metrics namely, accuracy [34], sensitivity and specificity. These metrics are elaborated in Table 1. In Table 1 $\mathrm{T}_{\mathrm{e}} \mathrm{P}_{\mathrm{s}}$ denotes true positive, $\mathrm{T}_{\mathrm{e}} \mathrm{N}_{\mathrm{g}}$ denotes true negative, $\mathrm{F}_{\mathrm{s}} \mathrm{P}_{\mathrm{s}}$ denotes false positive and $F_s N_g$ denotes false negative. $T_e P_s$ are the total AD individuals precisely detected into AD category, $\mathrm{T}_{\mathrm{e}} \mathrm{N}_{\mathrm{g}}$ are the total normal subjects precisely detected into normal category, $\mathrm{F}_{\mathrm{s}} \mathrm{P}_{\mathrm{s}}$ are the total normal subjects wrongly classified into AD category and $\mathrm{F}_{\mathrm{s}} \mathrm{N}_{\mathrm{g}}$ are the total AD subjects wrongly classified into normal category.

Table 1. Performance criteria

3.3 Performance comparison of different versions of SCELD with different wavelets for AD/N_oC, M_iCI/N_oC and AD/M_iCI classifications

Different versions of WSCELD have been tested in this work to evaluate the efficacy of the proposed method. The version DD-DTCWT+Energy-SCELD+PCA gives the best results among different versions. The reason is that DD-DTCWT shows higher directional selectivity than DD-DWT and DTCWT. The DD-DTCWT extracts features in sixteen directions while DD-DWT in eight directions and DTCWT in six directions. DD-DTCWT provides more efficient and directional features than DTCWT and DD-DWT.

The energy version of WSCELD outperforms other versions as it captures the energy contribution of the neighbouring pixels. Energy is the most promising parameter for finding structural changes in an image. The structural changes can be monitored through edge detection. Edges show the transformation between textured or smoothed regions and provide significant details about the position and morphological structure of pictured objects [35]. At edges, energy becomes maximum due to the rapid change in pixel values at its orthogonal direction. The ability of energy to capture edges lies in its emphasis on regions with high gradient magnitude or sharp transition in pixel values, indicating the presence of an edge. The squared gradient or energy of an image is an effective way to identify regions with abrupt changes in intensity, making it a key idea for edge detection [36, 37]. The other version like variance captures the local contrast information, the median captures the middle value of neighbouring pixels, and the mean captures the average value of neighbouring pixels. All these versions do not capture fine structural details thus leading to low performance. Table 2 indicates the performance of different versions of SCELD using a decision tree classifier with different wavelets. Table 3 indicates the p-value for the student t-test performed for the accuracy of different groups implemented in Table 2. The groups having p value greater than 0.05 are marked by *. These groups are H-V-ELBP/SCLD, H-V-ELBP/S-H-V-ELBP and SCELD/Mean-SCELD. They do not show significant improvement. However, the proposed version DD-DTCWT+Energy-SCELD+PCA gives outstanding results. The p-value obtained for the proposed version with other implemented versions is less than 0.05 in each case for AD classification. Energy SCELD is the best performer among variance, mean and median, whatever may be the wavelet and the best results are with DD-DTCWT. This shows the efficacy of the proposed method for early AD detection.

It can be observed from Table 2 that Circular Local Binary Pattern (CLBP) provides 84.2±1.4%, 83.2±1.4% and 81.5±1.2% classification accuracy for AD/N_oC, AD/M_iCI, and M_iCI/N_oC classifications respectively. These figures get improved by using Horizontal and Vertical Elliptical Local Binary Patterns (H-V-ELBP) because ELBP captures directional information also. The p-value obtained for group CLBP/H-V-ELBP is 2.60E-04 in Table 3 which is less than 0.05 and shows significant improvement for AD/N_oC classification. Shifted Circular Local Descriptors (SCLD) adds the shifted version in CLBP performance, so results get enhanced due to the contribution of different patterns. The p-value for the group CLBP/SCLD is 0.012 which shows significant improvement in SCLD. Shifted Circular Elliptical Local Descriptors (SCELD) further add the directional patterns thus improving the classification accuracy. Mean-SCELD, Median-SCELD, Energy-SCELD and Variance SCELD are the different versions of SCELD. The performance of Energy -SCELD is outstanding. The p-value for all groups from 7 to 10 in Table 3 is less than 0.05, which shows the significant improvement by Energy SCELD.

Table 2. Performance of different SCELD versions with different wavelets

Table 3. p-value for the t-test performed for the accuracy of different implemented versions in the paper

In all these models, histograms are used as features, so the number of textural features is 256 in each version. The model DD-DWT+SCELD+PCA combines SCELD with DD-DWT which gives 8 detail sub-bands on wavelet decomposition. SCELD is applied on each detail sub-band and subsequently each sub-band provides 256 textural features. This results in a total number of features of 256X8=2048. In the next step, Principal Component Analysis (PCA) is used to minimize the dimensionality problem. This model gives 91.2±1.2%, 88.3±1.3% and 86.1±1.7% for AD/N_oC, AD/M_iCI, and M_iCI/N_oC classifications respectively. Different versions of DD-DWT+SCELD like mean, median, variance and energy have been tested and the maximum results are obtained with the energy version. The p-value for all groups from 12 to 16 in Table 3 is less than 0.05 which shows the significant improvement by DD-DWT+Energy SCELD. The proposed methodology has been evaluated with three wavelets namely DD-DWT, DTCWT and DD-DTCWT. DTCWT gives 6 high-frequency sub-bands and DD-DTCWT gives 16 high-frequency sub-bands on wavelet decomposition. Thus, the number of histogram features with DTCWT and DD-DTCWT will be 256×6=1536 and 256×16=4096 respectively. Further PCA has been used to reduce dimensionality. Among all three wavelets, the best results are with DD-DTCWT and version DD-DTCWT+Energy-SCELD+PCA giving maximum results of 97.3±1.6%, 94.6±1.1% and 93.8±1.4% with AD/N_oC, AD/M_iCI, and M_iCI/N_oC classifications respectively. The p-value from groups 24 to 27 in Table 3 is less than 0.05 which shows the significance of improvement by DD-DTCWT+Energy-SCELD+PCA for AD classification at all stages. The Receiver Operating Characteristic (ROC) curve of different versions implemented in the paper is shown in Figure 8.

81.png

82.png

83.png

Figure 8. ROC plot for different implemented versions

Table 4. Performance of different classifiers

Results have been also checked with other classifiers like K-Nearest Neighbour (K-NN), Naive-Bayes (NB) and Linear Support Vector Machine (LSVM). The result of different classifiers with the proposed version is listed in Table 4 and in Figure 9.

9.png

Figure 9. Bar Plot of performance of different classifiers for the proposed version

3.4 Proposed and existing algorithms: Comparative analysis

The proposed method provides outstanding results in terms of accuracy, sensitivity, and specificity. These figures are 97.3±1.6%, 97.1±1.2% and 97.2±1.1%, 94.6±1.1%, 96.1±1.2% and 93.1±1.1% and 93.8±1.4%, 92.4±1.5% and 96.2±1.2% for AD/N_oC, AD/M_iCI, and M_iCI/N_oC classifications respectively. The proposed method has been compared with nine existing methods. The existing methods belong to the different types of AD detection techniques available in literature like wavelet transform-based techniques [38, 39], VBM techniques [12], ROI-based techniques [13, 14] and Texture-based techniques [19]. The proposed and existing methods have been executed ten times on MATLAB-19 using a 10-fold cross-validation technique and their average results are mentioned in Table 5.

The study [38] used DTCWT coefficients as features. This method of execution provides classification accuracy of 88.2±1.1%, 80.2±1.4% 82.0±1.6% for AD/N_oC, AD/M_iCI, and M_iCI/N_oC classifications respectively. The study [39] extracted statistical features by using the combination of DTCWT and its rotated version. This method of execution provides classification accuracy of 90.4±1.1%, 89.7±0.9% and 85.3±1.1% for AD/N_oC, AD/M_iCI, and M_iCI/N_oC classifications respectively. High dimensionality is the main limitation of studies [38, 39]. The study [12] used the VBM technique to select the appropriate grey density maps. A single atlas has been used in the study [12] which is mostly biased. This method of execution provides classification accuracy of 89.3±1.3%, 81.7±1.0% and 83.5±1.2% for AD/N_oC, AD/M_iCI, and M_iCI/N_oC classifications respectively. The studies [13] and [14] belong to ROI based technique as features are extracted from segmented hippocampus [13] and grey matter tissues [14] for AD classification. The study [13] on execution provides 85.3±1.3%, 74.2±1.8%, and 78.1±1.0% and the study [14] provides 88.5±1.2%, 86.2±1.6%, and 85.3±1.3% for AD/N_oC, AD/M_iCI, and M_iCI/N_oC classifications respectively. These studies [13, 14] need a high level of accuracy in segmentation. The recent study [40] is based on a deep learning approach and provides 86.7±1.2%, 83.5±1.6%, and 82.4±1.0% accuracy for AD/N_oC, AD/M_iCI, and M_iCI/N_oC classifications respectively. The p values obtained after performing t-tests on different groups of existing and proposed methods are illustrated in Table 6. Figure 10 indicates the ROC plot of proposed and existing algorithms for AD/N_oC, AD/M_iCI and M_iCI/N_oC classifications.

Table 5 states that the proposed method is 4.8% higher in accuracy for AD/N_oC classification, 3.3% for AD/M_iCI classification and 5.0% for M_iCI/N_oC classification compared to the best-performing existing method. The t-test has also been performed to check the significance of improvement. The p-value for all nine groups mentioned in Table 6 is less than 0.05 which shows the significance of improvement in the proposed technique in comparison to existing techniques.

The proposed technique extracts more informative features which can reflect both structural changes and grey matter density fluctuations. The combined strength of complex wavelet with SCELD helps in extracting structural details and grey matter density variations at a very minute level and thus, improves the classification accuracy. The proposed technique provides outstanding results for classifications at different stages of AD and gives more efficient results for early AD detection. The proposed technique is an automated technique for AD detection with more reliable results. Moreover, the proposed technique has been applied to the whole brain to capture the overall changes in the brain at different stages.

Table 5. Performance comparison of proposed and existing algorithms

Table 6. p-values for the t-test performed for the accuracy of the proposed method and existing methods

10.png

101.png

102.png

Figure 10. ROC plot of proposed and existing methods