Alzheimer's Disease Classification Using Distilled Multi-Residual Network

Published In

Applied Intelligence

Document Type


Publication Date



Early human intervention is crucial for diagnosing Alzheimer’s Disease (AD), since AD is irreversible and leads to progressive impairment of memory. In recent years, Convolutional Neural Networks (CNNs) have achieved dramatic breakthroughs in AD diagnosis. However, existing CNNs have difficulties in extracting subtle contextual information because of their structural limitations, i.e., it is difficult to extract discriminative features of several regions, such as hippocampus, parietal, temporal lobe tissues, and so on. In addition, current networks have difficulty in classifying imaging features with imbalanced data categories. Some loss functions can alleviate the above problems to some extent, but they are affected by outliers and trapped in local optimum easily. To address this issue, a Distilled Multi-Residual Network (DMRNet) is proposed for the early diagnosis of AD. The DMRNet consists of three main components: 1) Dense Connection Block, 2) Focus Attention Block, and 3) Multi-scale Fusion Module. The dense features are extracted by the Dense Connection Block while the local abnormal regions are refined by the Focus Attention Block and Multi-scale Fusion Module. Besides, to explore the hidden knowledge between each feature, a dilated classifier with self-distillation is proposed to ensemble several items pertaining to feature knowledge from feature space. Finally, the Remix Balance Sampler (RBS) is proposed to alleviate the influences of outliers. The proposed DMRNet is evaluated on baseline sMRI scans of the ADNI dataset. The result of experiments demonstrated that the proposed DMRNet not only achieves 7.15% greater accuracy than state-of-the-art methods but also successfully identifies some AD-related regions.


© 2022 Springer Nature Switzerland AG. Part of Springer Nature.



Persistent Identifier