NetFeatures-Transformer: Enhancing Brain Network Features to Classify Alzheimer's Disease

Yu Cao, Hongzhi Kuai, Ning Zhong, Bixin Wang, Fengmei Fan, Jianzhuo Yan

Abstract: The convergence of AI, big data, and magnetic resonance imaging (MRI) has unlocked trans formative potential in the field of neurological disease. For example, the resting-state fMRI technique is widely applied to identify functional alterations in brain and perform classification tasks for Alzheimer's Disease (AD). However, the high network complexity and difficult-to-interpret features of these models make it challenging to translate them into clinical applications. To overcome these limitations, the systematic Brain Informatics methodology is reconsidered to realize evidence combination and fusion computing from resting-state fMRI. In this work, a N etFeatures- Transformer is proposed to perform functional brain network classification through the integration of graph-based evidence, realizing interpretable discrimination of healthy controls (HC) and AD. The key contributions include: combining a transformer encoder with network nodal features to enhance classification performance, capturing different aspects of the brain network's topology. Additionally, the model utilizes the attention scores from trained transformer encoders to enhance interpretability, revealing the significant brain connections that have impact on the classification tasks at the same time.