DIPOG: A dynamic pooling graph neural network with spatial-temporal-frequency awareness for diagnosis of Alzheimer's disease.
Authors
Affiliations (7)
Affiliations (7)
- School of Communication and Information Engineering, Shanghai University, Shanghai, 200444, China.
- Department of Gerontology, Shanghai Sixth People's Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, Shanghai, 200233, China.
- Institute of Biomedical Engineering, School of Life Sciences, Shanghai University, Shanghai, 200444, China.
- Department of Neurology, Shanghai Pudong Hospital, Fudan University Pudong Medical Center, Shanghai, 201399, China.
- Department of Radiology, Beijing Friendship Hospital, Capital Medical University, Beijing, 100050, China.
- Department of Radiology, Beijing Friendship Hospital, Capital Medical University, Beijing, 100050, China. Electronic address: [email protected].
- Institute of Biomedical Engineering, School of Life Sciences, Shanghai University, Shanghai, 200444, China. Electronic address: [email protected].
Abstract
Resting-state functional magnetic resonance imaging (rsfMRI) discloses the spatiotemporal states of the brain and, through brain network modeling, enables the analysis of time-varying patterns associated with Alzheimer's disease (AD). To extract the spatial-temporal-frequency and dynamic features of brain networks and achieve generalizable AD diagnosis, this paper proposes a dynamic pooling graph neural network with spatial-temporal-frequency awareness. Specifically, a novel autoencoder is proposed to capture the accurate temporal features of rsfMRI through time interval representation learning. To improve the model's ability to fit abnormal spatial-temporal-frequency features associated with AD, a masked self-supervised learning strategy has been designed, and the reconstruction tasks are defined according to the masked brain regions and time periods. Considering the time-varying patterns in rsfMRI, a dynamic pooling method is employed, which utilizes dynamic time warping to merge the features of time steps exhibiting the same pattern, thereby avoiding the loss of time-varying features. A total of 2,857 samples from eight centers were included, and AD classification experiments were conducted. We compared 12 state-of-the-art models, and the results showed that the accuracy of our model was 6.4% higher than that of the best-performing comparison models. Furthermore, we used an explainability method to identify salient brain regions, and demonstrated that these regions were correlated with neuropsychological scale. In addition, AD progression could be reflected by the spatial-temporal-frequency features and time-varying patterns. This study presents a spatiotemporal learning framework that facilitates the individualized and precise diagnosis and treatment of AD.