MISS-Net: Multi-view Contrastive Transformer Network for MCI Stages Prediction Using Brain F-18-FDG PET Imaging

Mild Cognitive Impairment (MCI) is the transitional stage between healthy aging and dementia. MCI patients are characterized by very subtle changes in the brain. These changes with disease progression might assist the more precise dementia staging, which can reduce the number of Alzheimer's Disease (AD) patients through early intervention. Indeed, subjects diagnosed with MCI could be further divided into sub-categories (stable MCI and progressive MCI) and only part of them will convert to dementia. In this paper, we propose a multi-view contrastive transformer network for MCI sub-categories detection with the aim of early AD conversion prediction. The proposed method is based on a two-stage learning scheme that optimally captures local and global information from F-18 FluoroDeoxyGlucose Positron Emission Tomography (F-18-FDG PET) images. The proposed approach optimally exploits the complementary of the three image projections (axial, sagittal, and coronal), through contrastive learning, for efficient multi-view clinical pattern (embedding) learning. The proposed method has been evaluated on a subset of the ADNI dataset. Obtained results outperform recent uni-modal and multi-modal state-of-the-art approaches in (sMCI) vs. (pMCI) detection. We report an average accuracy, sensitivity, and sensitivity of respectively 87.13%, 90.61%, and 83.65%.