MTC-MSFFNet: a multi-task classification model based on multi-source feature fusion for Alzheimer's disease

Predicting the stages of Alzheimer's disease (AD) is crucial for delaying disease progression and enabling early intervention. A large amount of existing research focuses on the classification of cognitively normal (CN), mild cognitive impairment (MCI), and AD. However, the two subtypes of MCI-stable mild cognitive impairment (sMCI) and progressive mild cognitive impairment (pMCI)-should not be overlooked. Therefore, this study aims to accurately diagnose the disease stage of patients (CN, MCI, or AD) and further distinguish between sMCI and pMCI. In this work, a multi-task classification model based on multi-source feature fusion, termedMTC-MSFFNet, is proposed to accomplish two diagnostic tasks: (1) CN versus MCI versus AD, and (2) sMCI versus pMCI. We select the hippocampus (HIP) and entorhinal cortex (ERC) as feature maps for the three-class task, and the hippocampus (HIP) and gray matter (GM) for the sMCI/pMCI task. The MTC-MSFFNet integrates a multi-source feature fusion module (combining brain structure maps with structural magnetic resonance imaging (sMRI) data), a task-specific weight learning module guided by brain structural information, and dedicated task heads for each classification objective. The proposed method is evaluated on a mixed dataset constructed from the Alzheimer's Disease Neuroimaging Initiative (ADNI) and the Open Access Series of Imaging Studies (OASIS). Experimental results demonstrate that MTC-MSFFNet achieves an average accuracy of 98.09% for CN versus MCI versus AD classification and 95.16% for sMCI versus pMCI classification. These results indicate that the proposed approach has significant potential to assist clinicians in developing targeted and personalized treatment plans.