3D convolutional neural networks with hybrid attention mechanism for early diagnosis of Alzheimer’s disease

Abstract

As a non-invasive and radiation-free imaging technique, magnetic resonance imaging (MRI) can intuitively display the three-dimensional tissues and structures of human brain, showing the great prospect in the early screening and diagnosis of Alzheimer’s disease (AD). MR image processing on the basis of deep learning methods has aroused increasing attention, and the core of this type of method is to construct an efficient model to recognize and extract the key features of the images. In this article, a 3D Residual U-Net model incorporating hybrid attention mechanism (3D HA-ResUNet) is proposed for the auxiliary diagnosis of AD using 3D MR images. The backbone classification model consists of an up-sampling branch network, a down-sampling branch network, and intermediate connection residual blocks. The hybrid attention mechanism exploits the advantages of both channel and spatial attention, and is merged with the skip connection of the backbone classification model. In the binary classification task of AD vs. normal cohort (NC) on the ADNI dataset, the addition of the hybrid attention module helps improve accuracy, sensitivity, precision, F1 score and G-mean by 4.88%, 10.52%, 0.94%, 6.17% and 5.60%, respectively. Furthermore, the proposed method demonstrates superior generalization ability compared with other state-of-the-art methods. The 3D HA-ResUNet was further tested in the mild cognitive impairment (MCI) subtype classification task on the local dataset and achieved 100% of accuracy. In addition, an attribution-based visual interpretability method is employed to reveal the regions and features that the proposed model focuses on for classification. The visual interpretations combined with domain knowledge are capable of providing a valuable reference for physicians’ clinical decision-making.