关键词: 分形特征, 连接比例, 脑网络, 机器学习, 疾病诊断

Abstract: Most existing studies construct low-resolution functional brain networks for Alzheimer’s disease (AD) using the AAL90 atlas to examine its impact on network topology. However, high-resolution brain networks can more accurately capture subtle alterations caused by AD. In contrast, classification based on a single-resolution network often results in lower diagnostic accuracy. To address this limitation, we constructed brain networks at three resolutions (BN246, AAL625, and AAL1024) for cognitively normal (CN) individuals and patients with AD, and analyzed their fractal features. Meanwhile, we developed a classification framework for AD that fuses fractal features from multiple brain atlases. By fusing the fractal properties of networks constructed from the three atlases, we derived two novel feature vectors: the multi-atlas fused connection ratio (MAFCR) and the multi-atlas fused fractal dimension (MAFFD). Experimental results showed that brain networks at all three resolutions exhibited fractal properties in both CN and AD groups, with marked differences in fractal dimension and average connection ratio between them. Compared to the average connection ratio from single-resolution networks, the MAFCR achieved a diagnostic accuracy of 84.74% and an F1 score of 84.53%, representing improvements of approximately 18% and 16%, respectively. Furthermore, the MAFCR outperformed multi-atlas fused topological features such as average degree, clustering coefficient, and characteristic path length by approximately 12% in accuracy and 13% in F1 score.

Key words: fractal feature, connection ratio, brain network, machine learning, disease diagnosis