EARLY STAGE ALZHEIMER’S DISEASE DETECTION USING WAVELET BASED APPROACHES

Abstract

Alzheimer's disease (AD) is a neurodegenerative disorder characterized by progressive cognitive decline. Early and accurate AD detection is crucial for better patient outcomes, effective intervention, and care planning.AD is not completely curable at present, but its progression can be slowed down by its early and timely detection. Over the decade, many efforts have been made to develop algorithms for early AD detection, but it is a challenging task due to the overlapping of AD symptoms with normal aging, lack of definitive biomarkers, AD heterogeneity, and dearth of techniques for asymptomatic AD detection. To mitigate these challenges, there is a need to develop feature extraction techniques that can detect subtle structural changes in the AD brain with fluctuations in grey matter density without having the limitations of existing feature extraction techniques. Existing techniques generally suffer from the problem of high dimensionality, imperfect registration of images over template, inaccurate segmentation of specific brain regions, and lack of standard protocols for the acquisition of neuroimaging data. This thesis focuses on developing a novel framework for early AD detection using neuroimaging data and advanced feature extraction techniques. This research utilizes Magnetic Resonance Imaging (MRI) datasets and introduces feature extraction methods such as direct coefficient-based feature extraction, histogram-based feature extraction, statistical features-based feature extraction, and mean energy-based feature extraction techniques for AD classification from Normal Controls (NC) and Mild Cognitive Impairment (MCI). All the proposed techniques employed complex wavelet transform to decompose the MR image into sub-bands, and further features are extracted from the detail sub-bands in different ways. The detail sub-bands are high-frequency sub-bands and capture high-frequency variations in grey matter density due to AD. Complex wavelets are also good at capturing structural changes in the brain due to AD. This thesis comprises a total of seven proposed methods in four chapters to address the above-mentioned problems. In the first chapter, a brief introduction of Alzheimer’s Disease with its biomarkers and neuro-imaging tests have been discussed. An intense literature ix review is done in the second chapter, which includes the existing state-of-the-art methods for feature extraction and classification for early AD detection. The first method, i.e., Direct coefficient-based feature extraction technique using multiple 2-D MR slices, is proposed in the third chapter. Multiple brain slices are more effective compared to a single slice when disease spreads in different parts of the brain. In this technique, a 3-D MR image is divided into 32 2-D slices of size 256 x 256, and then a Double Density-Dual Tree Complex Wavelet Transform (DD-DTCWT) is applied on the 32 2-D MR center slices, and fifth-level detail coefficients (2048 coefficients) are extracted. Each 3-D MRI provides 32 x 2048 coefficients (65536 coefficients). These coefficients are directly used as features. The high dimensionality of features is reduced by using Principal Component Analysis (PCA) in the second stage. At the third stage, a feature selection algorithm, i.e., Genetic Algorithm (GA), is introduced to improve the classification results. The proposed method is effective in detecting the different stages of AD as it captures both grey matter density fluctuations and structural variations in the brain with the advancement of the disease. This work also investigates the performance of conventional statistical methods, namely, t- test and evolutionary algorithms viz. Particle Swarm Optimization (PSO), Ant Colony Optimization (ACO), and GA for feature selection in AD classification. Two different novel methods are proposed for early AD detection in the fourth chapter. These methods are histogram-based feature extraction techniques. In these methods, first level DD-DTCWT has been applied on 2-D MR images, and sixteen sub-bands have been obtained. Now, in the first method, Edge Weighted Local Binary Pattern (EWLBP) has been applied on sixteen sub-bands, and histograms are extracted. These histograms are used for early AD detection. EWLBP consists of Sobel operator which incorporates the edge information while computing the binary patterns. The centre pixel, which is having the highest gradient magnitude, is given more weightage while generating EWLBP codes. In the second method, Wavelet-based Shifted Circular Elliptical Local Descriptor (WSCELD) is applied in place of EWLBP, and histogram features are extracted. These histograms are used for early AD detection. WSCELD includes the properties of circular and elliptical LBP and thus can detect isotropic and anisotropic structural details from the images. The shifted version of CELD helps in extracting multiple micro and macro patterns, which indicates the atrophies at cellular level as well as at large scale. In the second method, the performance of different versions, like Mean WSCELD, Median WSCELD, Energy WSCELD, and Variance WSCELD, has been investigated for histogram features. Both techniques are efficient for early AD detection x and operate with lower dimensionality, making it more efficient in comparison to the direct coefficient-based feature extraction technique. Three different novel methods are proposed for addressing the issue of scalability in the fourth chapter. These techniques also provide excellent results for early AD detection on two datasets and with multiclass classification. These are statistical feature extraction techniques. Statistical features are less sensitive to noise as they focus on the overall data distribution rather than individual bins in histograms. All three methods, 2-D MR images are decomposed into sixteen sub-bands by applying first level DD-DTCWT. In the first method, a shifted Elliptical Local Binary Pattern (ELBP) has been applied on sixteen sub-bands, and multiple micro and macro patterns have been obtained from these sub-bands. These patterns reflect the different directional texture patterns in the images. Six statistical features, like mean, median, energy, kurtosis, variance, and skewness have been obtained from these patterns and used for AD classification. Eight-neighbour shifted ELBP is used, and analysis with eight different patterns is included in the chapter. This method shows that the concatenation of features from all eight patterns enhances the classification accuracy in comparison to the features contributed by a single pattern. The total number of statistical features in this method is 768. In the second method, the Grey Level Co-occurrence Matrix (GLCM) has been computed for sixteen sub-bands, and 22 statistical features have been obtained from these matrices. The total number of features is 352. A Genetic Algorithm has been used for selecting the appropriate features. These statistical features have been used for AD classification. In the third method, Shifted Circular Elliptical Local Descriptor (SCELD) has been applied to sixteen sub-bands, and multiple micro and macro patterns have been obtained from these sub- bands. Statistical features have been obtained from these patterns and used for AD classification. SCELD includes the properties of circular and elliptical LBP and can extract isotropic and anisotropic structural details from the images. A novel method using mean energy-based feature extraction technique is proposed in the sixth chapter. It can address the problem of spatial information loss, which results when feature dimensionality reduction techniques are applied to high dimensional data. In addition to this, the proposed technique provides novel features for accurate and early AD detection without using any feature dimensional reduction techniques. The results with this technique are outstanding for all AD/NC, AD/MCI, and MCI/NC classifications and multiclass classification. Thus, this technique is also scalable. This technique is quite simple but more efficient in comparison to the previous techniques as this technique does not require any xi dimensionality reduction or feature selection technique. In this technique, MR images are decomposed up to the fifth level using DD-DTCWT, and complex wavelet coefficients from all sixteen sub-bands are extracted at each level. DD-DTCWT has sixteen sub-bands at each level of decomposition, so up to the fifth level, 80 complex sub-bands are there. The proposed technique introduces a novel feature, i.e., the mean energy of real and imaginary components taken independently for early AD detection. This feature efficiently captures the grey matter density and structural abnormalities in the brain and retains the ability to detect AD in its earliest stage, i.e., the MCI stage. This feature batch includes the mean energy of real and imaginary sub-bands from the first level to fifth level. Mean energy of real sub-bands provides the average information, and the mean energy of imaginary sub-bands provides the structural details of the image. The total number of features in this batch is 160, which is quite low in dimensionality in comparison to previous techniques. The performance of other features, like mean energy of real coefficients, mean energy of imaginary coefficients, mean energy of complex coefficients, variance of the magnitude of complex coefficients, and entropy of complex coefficients, are also investigated in the proposed technique. Finally, in this thesis, the conclusions inferred from this research work are summarized, and potential future work in this area is highlighted.