Deep joint learning diagnosis of Alzheimer’s disease based on multimodal feature fusion

Abstract

Abstract Alzheimer’s disease (AD) is an advanced and incurable neurodegenerative disease. Genetic variations are intrinsic etiological factors contributing to the abnormal expression of brain function and structure in AD patients. A new multimodal feature fusion called “magnetic resonance imaging (MRI)-p value” was proposed to construct 3D fusion images by introducing genes as a priori knowledge. Moreover, a new deep joint learning diagnostic model was constructed to fully learn images features. One branch trained a residual network (ResNet) to learn the features of local pathological regions. The other branch learned the position information of brain regions with different changes in the different categories of subjects’ brains by introducing attention convolution, and then obtained the discriminative probability information from locations via convolution and global average pooling. The feature and position information of the two branches were linearly interacted to acquire the diagnostic basis for classifying the different categories of subjects. The diagnoses of AD and health control (HC), AD and mild cognitive impairment (MCI), HC and MCI were performed with data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). The results showed that the proposed method achieved optimal results in AD-related diagnosis. The classification accuracy (ACC) and area under the curve (AUC) of the three experimental groups were 93.44% and 96.67%, 89.06% and 92%, and 84% and 81.84%, respectively. Moreover, a total of six novel genes were found to be significantly associated with AD, namely NTM, MAML2, NAALADL2, FHIT, TMEM132D and PCSK5, which provided new targets for the potential treatment of neurodegenerative diseases.

Keywords

• Alzheimer’s disease
• Multimodal feature fusion
• Deep joint learning diagnosis
• Attention mechanism
• ResNet