アルツハイマー病における神経心理学的および神経画像データの統合による軽度認知障害の特定能力向上に向けたビジョン・言語モデル

Alzheimer's Disease (AD) is a prevalent neurodegenerative disorder with the progression typically spanning from mild cognitive impairment (MCI) to severe dementia. However, in clinical practice, diagnosing MCI is challenging due to conflicts between neuroimaging findings and clinical neuropsychological assessments. Structural and metabolic changes in imaging may not immediately correlate with cognitive symptoms, making early MCI detection difficult. This diagnostic complexity requires considerable clinical expertise and can lead to delays in detection. Moreover, multi-modal fusion using computer-based techniques faces challenges due to the inherent heterogeneity across different data modalities. In this study, a Vision-Language Model (VLM)-based approach is proposed to enhance the identification of Alzheimer's Disease, with a particular focus on the early detection of MCI. Our contribution is a synergistic architecture that includes domain-specific learnable abnormality tokens that function as adaptive probes for clinical pathologies, and a unique Unified Multi-Modal Attention module designed to explicitly harmonize conflicting signals between neuroimaging and clinical data. Finally, a Large Language Model synthesizes all information to generate the final diagnostic output. Evaluated on three publicly available datasets across four identification tasks, the proposed method performs robustly, outperforms state-of-the-art approaches, especially in the challenging MCI identification task, and significantly reduces diagnostic errors when dealing with conflicting data. The results underscore the practical significance of integrating multi-modal data within a VLM framework, which not only enhances diagnostic accuracy but also reduces diagnostic errors and delays in early-stage MCI identification, thereby supporting more timely and effective clinical decision-making.