基于HubGLasso注意力机制的脑网络分类研究

Research on Brain Network Classification Based on HubGLasso Attention Mechanism

脑网络分类有助于脑疾病的早期诊断,也有益于理解脑疾病发病机理,具有重要的研究与应用价值。其中,卷积神经网络应用广泛,可以提取脑网络的拓扑特征,是脑网络分类中的一个前沿热点。然而,现有方法未考虑脑网络中Hub节点对脑功能的重要贡献,这可能会导致特征提取不充分,限制了它们的分类性能。为此,该文提出了一种基于HubGLasso注意力机制的卷积神经网络模型,用于进行脑网络分类任务。该方法包含了一种新的卷积层结构,首先利用GLasso模型去除脑网络中的冗余信息,然后引入Hub约束与注意力机制,使其能够提取与异常Hub结构相关的重要特征,并用于脑疾病诊断。实验结果表明,该方法在包含1112个被试的真实自闭症数据集上取得了68.67%的准确率,显著优于目前已有方法,证明了其应用价值。更进一步,通过对训练后的模型进行特征分析,能够得到与脑疾病相关的脑区信息与Hub节点结构信息,为脑疾病病理机制的研究提供了全新的视角。

Brain network classification is useful for early diagnosis of brain diseases and understanding the pathogenesis of brain diseases,which has important research and application values.Among them,convolutional neural network is widely used and can extract the topological features of brain network,which is a frontier hot spot in brain network classification.However,existing methods do not consider the important contribution of Hub nodes in brain networks to brain function,which may lead to inadequate feature extraction,limiting their classification performance.To this end,we propose a convolutional neural network model based on the HubGLasso attention mechanism for brain network classification tasks.It contains a new convolutional layer structure,which first removes redundant information from the brain network using the GLasso model,and then introduces Hub constraints and attention mechanisms that enable it to extract important features associated with abnormal Hub structures and use them for brain disease diagnosis.The experiments showed that the proposed method achieved 68.67% accuracy on a real autism dataset containing 1112 subjects,which was significantly better than the existing methods and proved its application value.Further,the trained model can be characterized to obtain the information of brain regions and Hub node structures related to brain diseases,which provides a new perspective on the pathological mechanism of brain diseases.