注意力密集连接网络的早期AD脑形态学表征与分类
Brain morphological representation and classification of early AD based on attention densely connected network
康迪 1赵敏 1程和伟 1田银 1王伟 1李章勇1
作者信息
- 1. 重庆邮电大学 生命健康信息科学与工程学院,重庆 400065
- 折叠
摘要
针对阿尔茨海默病(Alzheimer's disease,AD)这种神经退行性疾病早期阶段难以被及时发现,无法进行诊断与干预的问题,提出基于注意力密集连接网络的早期AD脑形态学特征提取与分类方法.采用密集连接网络为主干架构、多视角三维图像信息作为网络输入的设计策略,引入注意力机制使得网络能够捕获对AD分类具有重要贡献的脑区.实验结果表明,提出算法对认知正常(cognitively normal,CN)与轻度认知障碍(mild cognitive impair-ment,MCI)、MCI与AD、CN与AD的分类正确率依次达到了 98.37%、97.63%、98.60%,在AD分类领域处于较高水平.此外,通过对注意力机制得到的注意力图进行分析,可发现AD患者脑形态学演变轨迹,由CN转化为MCI涉及皮层下结构的脑形态学异常改变,再转化为AD则进一步涉及皮层结构的脑形态学异常改变.
Abstract
To address the challenges of timely detection,diagnosis,and intervention in the early stages of Alzheimer's dis-ease(AD),a neurodegenerative disorder,we propose a method for extracting and classifying early AD brain morphological features based on attention densely connected network.The design strategy of using a dense connection network as the back-bone architecture and multi-view 3D image information as the network input is adopted,and an attention mechanism is in-troduced to enable the network to capture brain regions that are important for AD classification.Experimental results show that under the experimental environment of this study,the classification accuracy rates for cognitively normal(CN)vs.mild cognitive impairment(MCI),MCI vs.AD,and CN vs.AD have reached 98.37%,97.63%,and 98.60%respectively,re-presenting an advanced level in the field of AD classification.Moreover,through the analysis of attention maps obtained via the attention mechanism,we can identify the evolution trajectory of AD brain morphology.The transformation from CN to MCI involves abnormal changes in brain morphology in subcortical structures,and the further transformation to AD involves abnormal changes in brain morphology in cortical structures.
关键词
阿尔茨海默病/脑形态学/密集连接网络/注意力机制/分类Key words
Alzheimer's disease/brain morphology/densely connected network/attention mechanism/classification引用本文复制引用
基金项目
重庆市留学人员回国创业创新支持计划项目(2007010003947888)
出版年
2024
国家科技期刊平台
NSTL
万方数据