Attention-Guided Deep Domain Adaptation for Brain Dementia Identification with Multi-site Neuroimaging Data

Abstract

Deep learning has demonstrated its superiority in automated identification of brain dementia based on neuroimaging data, such as structural MRIs. Previous methods typically assume that multi-site data are sampled from the same distribution. Such an assumption may not hold in practice due to the data heterogeneity caused by different scanning parameters and subject populations in multiple imaging sites. Even though several deep domain adaptation methods have been proposed to mitigate data heterogeneity between sites, they usually require a portion of labeled target data for model training, and rarely consider the potentially different contributions of different brain regions to disease prognosis. To address these limitations, we propose an attention-guided deep domain adaptation (AD$^2$A) framework for brain dementia prognosis, which does not need label information of the target domain and can automatically identify discriminative locations in whole-brain MR images. The proposed AD$^2$A framework consists of three key components: 1) a feature encoding module for representation learning of input MR images, 2) an attention discovery module for automatically locating dementia-related discriminative regions in brain MRIs, and 3) a domain transfer module with adversarial learning for knowledge transfer between the source and target domains. Extensive experiments have been conducted on three benchmark neuroimaging datasets, with results suggesting the effectiveness of our method in both tasks of brain dementia identification and disease progression prediction.