Abstract:

Magnetic resonance imaging and positron emission tomography are imaging techniques commonly used for the early diagnosis of Alzheimer's disease. Combining these two modalities enables a more comprehensive assessment of brain status using both anatomical and metabolic information. However, traditional multimodal fusion is mainly through channel splicing, which fails to fully utilize the complementary information between modalities and affects the effectiveness of the model. To this end, this paper proposes a multi-task Alzheimer's disease classification model based on adversarial learning and cross-attention. The model reduces the feature differences between different modalities through adversarial learning and uses cross-attention for feature fusion, while the brain age prediction task is used as an auxiliary task to enhance the classification performance. The experimental results show that the method achieves an accuracy rate of 91.10% and an F1 score of 91.01% in the classification tasks of Alzheimer's disease, mild cognitive impairment and normal control classification tasks, respectively, which not only improves the accuracy of early diagnosis, but also enhances the ability to monitor the disease progression, and provides an important support for clinical interventions in Alzheimer's disease.

Key words: adversarial learning, cross-attention mechanisms, multitasking, Alzheimer's disease, multimodality