文章摘要

文章信息/Info

Title:: Alzheimer′s disease classification based on residual self attention mechanism

作者:: 卢星进1; 2; 王书卜3; 肖磊1; 2△; 高礼彬1; 2; 李瑞1; 2; 胡众义1; 2△; （1.温州大学计算机与人工智能学院，温州 325035；2.温州市智能影像处理与分析重点实验室，温州 325035；3.上海东方医院，上海 200120）

Author(s):: LU Xingjin1; 2; WANG Shubu3; XIAO Lei1; 2; GAO Libin1; 2; LI Rui1; 2; HU Zhongyi1; 2; （1. College of Computer Science and Artificial Intelligence,Wenzhou University,Wenzhou 325035,China；2.Key Laboratory of Intelligent ImageProcessing and Analysis, Wenzhou 325035；3.Shanghai East Hospital, Shanghai 200120,China）

Keywords:: Alzheimer′s disease; Classification; Structural magnetic resonance imaging; Residual neural networks; Self-Attention mechanism; Adaptive gradient cropping

摘要:: 针对基于磁共振成像（magnetic resonance imaging, MRI）切片的阿尔茨海默症（Alzheimer′s disease, AD）诊断研究方法存在切片划分权重分配不当,导致模型弱化核心切片区间，降低核心区间特征信息比重等问题。本研究提出一种基于残差自注意力机制框架以实现AD的准确分类，模型包括残差模块和自注意力机制。残差模块用于学习并提取每张图像的特征信息；自注意力机制用于学习切片间的特征信息，动态分配切片权重，增强核心分类区间比重，最终集成所有权重结果进行AD分类。经验证，本研究算法在AD和健康对照分类中的准确率（ACC）、召回率（REC）和均衡平均数（F1-Score）相较于基准方法分别提高了2.4%、4.3%、1.4%。该方法可精确、高效地区分阿尔茨海默症和健康对照组的MRI图像，能够有效地辅助医生实现AD的准确诊断。

Abstract:: Magnetic resonance imaging (MRI) slice based Alzheimer′s disease (AD) diagnostic study method has inappropriate slice division weight assignment, which leads to the model weakening the core slice interval and reducing the core interval feature information. We proposed a residual-based self-assignment model, which mainly included a residual module and a self-attentive mechanism. The residual module was used to learn and extract the feature information of each slice image. The self-attention mechanism was used to learn the feature information between slices, assign the slice weights dynamically, enhance the core classification interval weight, and finally integrated all the weight results for AD classification. The accuracy (ACC), recall (REC), and equilibrium mean (F1-Score) of this method in the classification of AD and healthy controls increased 2.4%，4.3%，1.4%，respectively，comparing with benchmark method, this method can accurately and efficiently distinguish between MRI images of Alzheimer′s disease and healthy controls，and can effectively assist doctors to realize the accurate diagnosis of AD。