Efficient Alzheimer's Disease Classification Based on AlexNet Model

Alzheimer's disease (AD) generally results in irreversible brain damages. Early diagnosis of disease plays an important role in preventing the progression of AD. Deep convolutional neural networks (CNN) have achieved prominent performance in the field of natural image recognition, while some problems exist in applying a classic CNN model on 3D MRI for AD classification. To address these issues, with 194 AD subjects, late mild cognitive impairment (LMCI) 123 subjects and 105 normal control(NC) subjects, a hybrid computational strategy was proposed based on a pre-trained AlexNet CNN model and sMRI for AD classification. The feature presentation of pre-trained network was efficiently transferred to AD classification task by using transfer learning, 3D features reconstruction, feature reduction using Max pooling and principal component analysis (PCA), and selection feature using sequential forward search and (SFS) method. Then, support vector machines (SVM) was applied to classification. The classification accuracy values on conv3, conv4, conv5 layers of AlexNet were 89.93%, 91.28%, and 87.25% of AD/NC, respectively, 80.77%, 76.92%, and 78.21% of AD/LMCI, respectively, 72.46%, 75.45%, and 73.65% of NC/LMCI, respectively. Results show that features extracted from a classic CNN model and their 3D reconstruction can achieve good performance on AD classification.