Detection of Alzheimer's disease and mild cognitive impairment based on structural volumetric MR images using 3D-DWT and WTA-KSVM trained by PSOTVAC

• Published in: Biomedical signal processing and control Vol. 21; pp. 58 - 73
• Main Authors: Zhang, Yudong, Wang, Shuihua, Phillips, Preetha, Dong, Zhengchao, Ji, Genlin, Yang, Jiquan
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.08.2015
• Subjects: Kernel SVM; Magnetic resonance imaging; Multiclass SVM; Particle swarm optimization; Time-varying acceleration-coefficient; SES; MR(I); RS; TVAC; WTA; ALS; (M)(K)SVM; KNN; nWBV; SA; 1D/2D/3D-DWT; DS; CDR; eTIV; OASIS; ASF; Kernel SVM; MMSE; RBF; TN/FN(R); (D)(F)WT; GA; PC(A); CA; MSE; QP; AD; SVMDT; Time-varying acceleration-coefficient; CM; TP/FP(R); DAG; Particle swarm optimization; MWV; PSO; ROI; MCI; CV; Magnetic resonance imaging; Multiclass SVM; SMO; NC
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2015.05.014

•We take the whole brain (not the ROIs) as the research objective, so there is no need for brain segmentation.•The sensitivity of NC is up to 93.81%. The specificities of MCI and AD are 93.39% and 92.21%.•We use 3D-DWT to capture the 3D texture feature of brain, use ALS-PCA for feature reduction of dataset containing missing attributes.•We use TVAC-PSO to get the optimal kernel parameter of each individual KSVM.•We compare three different multiclass KSVM methods, and find that WTA performs best. We proposed a novel classification system to distinguish among elderly subjects with Alzheimer's disease (AD), mild cognitive impairment (MCI), and normal controls (NC), based on 3D magnetic resonance imaging (MRI) scanning. The method employed 3D data of 178 subjects consisting of 97 NCs, 57 MCIs, and 24 ADs. First, all these 3D MR images were preprocessed with atlas-registered normalization to form an averaged volumetric image. Then, 3D discrete wavelet transform (3D-DWT) was used to extract wavelet coefficients the volumetric image. The triplets (energy, variance, and Shannon entropy) of all subbands coefficients of 3D-DWT were obtained as feature vector. Afterwards, principle component analysis (PCA) was applied for feature reduction. On the basic of the reduced features, we proposed nine classification methods: three individual classifiers as linear SVM, kernel SVM, and kernel SVM trained by PSO with time-varying acceleration-coefficient (PSOTVAC), with three multiclass methods as Winner-Takes-All (WTA), Max-Wins-Voting, and Directed Acyclic Graph. The 5-fold cross validation results showed that the “WTA-KSVM+PSOTVAC” performed best over the OASIS benchmark dataset, with overall accuracy of 81.5% among all proposed nine classifiers. Moreover, the method “WTA-KSVM+PSOTVAC” exceeded significantly existing state-of-the-art methods (accuracies of which were less than or equal to 74.0%). We validate the effectiveness of 3D-DWT. The proposed approach has the potential to assist in early diagnosis of ADs and MCIs.