Multi-classification of Alzheimer's Disease by NSGA-II Slices Optimization and Fusion Deep Learning

This paper presents a two-phase hierarchical classifier for determining the different states in Alzheimer's disease (AD). In the first phase, an evolutionary system is developed to determine the most relevant slices (in both X-axis and Y-axis) of the magnetic resonance imaging (MRI) for the construction of a classifier. To obtain the image features, the biorthogonal wavelet transform 3.3 was used at level 2. Due to the high number of coefficients, a dimensionality reduction is performed using minimum Redundancy - Maximum Relevance algorithm (mRMR) and Principal Component Analysis (PCA). An evolutionary algorithm on a high-performance computer with GPU was used to optimize the slides. Support vector machine (SVM) was used in the fitness function to estimate the features of the classifier in a computationally simple way. In the second phase, using the different solutions of the Pareto front obtained by the evolutionary algorithm, a multiple deep learning system was developed, each of the systems having as input one of the selected slices of the analyzed solution. The solution with three slices (trade-off between complexity and accuracy) was used as the solution. The obtained hierarchical deep learning system fused the information from each system and analyzed the probabilities obtained for each class. As a final result, an accuracy of 92% was obtained for the six classes. A total of 1,200 patients from the Alzheimer's disease neuroimaging initiative (ADNI) database were used, corresponding to six different classes of patients (with varying degrees of dementia).