DSC-CNN: A Dual-Stream CNN with Cognitive Embedding Fusion for Early Alzheimer’s Diagnosis

Alzheimer’s disease (AD) remains one of the most prevalent and challenging neurodegenerative disorders, with early diagnosis being crucial for timely intervention. In this paper, a novel dual-stream deep learning architecture, termed DSC-CNN (Dual-Stream CNN with Cognitive Embedding Fusion), has been proposed to enhance the accuracy and interpretability of AD classification. The model integrates volumetric MRI data with structured clinical metadata through two dedicated processing streams: a spatial ResNet3D-18 backbone with attention for anatomical features and a lightweight encoder for cognitive attributes. These complementary embeddings have been fused via a bilinear attention mechanism, allowing the model to capture intricate cross-modal interactions. To ensure both generalizability and transparency, the framework has incorporated intrinsic attention visualization and prototype-guided decision paths in place of traditional post-hoc explanation tools. Experiments have been conducted on the ADNI and OASIS datasets, demonstrating that the proposed DSC-CNN has achieved a classification accuracy exceeding 99.68%, outperforming several recent related methods. The model has shown particular strength in identifying early mild cognitive impairment (EMCI) cases while maintaining a compact parameter footprint, enabling efficient deployment in clinical settings. These results suggest that DSC-CNN is a robust, interpretable, and scalable solution for improving AD diagnosis.