Attention-based and micro designed EfficientNetB2 for diagnosis of Alzheimer’s disease

Abstract

Recently, many deep learning methods have been successfully used to diagnose Alzheimer’s disease (AD) using brain imaging. However, structural magnetic resonance imaging (sMRI) of AD detects relatively small lesion areas, and it is difficult to distinguish the early lesions. These factors make it difficult to extract the key features in the subtle regions for discrimination in many studies. To address these issues, an attention-based and micro designed EfficientNetB2 (EB2) approach is proposed in this study to classify AD, mild cognitive impairment (MCI), and normal control (NC). First, the front portion of the EB2 uses a global attention mechanism (GAM) to improve the classification results by capturing significant features in three dimensions. Next, coordination attention (CA) was added to the EB2 model. The CA can automatically extract channel and location information from the sMRI two-dimensional slice data. The location information is critical for constructing spatial attention maps that help the model to identify the lesion areas accurately, thereby helping the model to extract features that are useful for classification. Finally, the model uses the micro design (MD) method of the ConvNeXt network, which investigates the effect of activation function and batch normalization layer on the model at the microscopic scale. MD can reduce the model complexity while improving classification ability considerably. The accuracy of the proposed method is 93.30%, 92.42% and 92.03% for AD/NC, AD/MCI and MCI/NC dichotomous data, respectively. Finally, the proposed method outperforms the existing convolutional neural networks, such as AlexNet, GoogleNet, MobileNetV2, MobileNetV3, and DenseNet.