A modified 3D EfficientNet for the classification of Alzheimer's disease using structural magnetic resonance images

Abstract

Structural magnetic resonance imaging (MRI) provides useful information for biomarker exploration and intelligent classification of Alzheimer's disease (AD). Machine learning and deep learning methods have been used in feature extraction and computer-aided diagnosis or prediction of AD. In this research, a new deep learning model is developed to detect or predict AD in an effective manner. A modified 3D EfficientNet with a sequentially connected 3D mobile inverted bottleneck convolution (MBConv) block is proposed to explore the multiscale characteristics of brain MR images for AD classification. The 3D MBConv block consists of depthwise convolution and a squeeze-and-excitation module to transform the input features into more compact features with fewer parameters than those in standard convolution. The proposed method was evaluated considering four classification tasks: (1) NC versus AD, (2) NC versus all MCI, (3) NC versus pMCI, and (4) sMCI versus pMCI. The proposed model achieved accuracies of 95.00%, 86.67%, and 83.33% for the classification of NC versus AD, NC versus pMCI, and sMCI versus pMCI, respectively, and exhibited a high performance in comparison with the classical networks and several existing methods. Efficient deep learning networks are powerful tools for the early diagnosis and prediction of AD.