迁移学习特征提取的rs-fMRI早期轻度认知障碍分类

PDF下载HTML

Title:

Transfer learning-based feature extraction method for the classification of rs-fMRI early mild cognitive impairment

作者:

孔伶旭¹, 吴海锋^1,2, 曾玉^1,2, 陆小玲¹, 罗金玲¹

1. 云南民族大学 电气信息工程学院，云南 昆明 650500;
2. 云南省高校智能传感网络及信息系统创新团队，云南 昆明 650500

Author(s):

KONG Lingxu¹, WU Haifeng^1,2, ZENG Yu^1,2, LU Xiaoling¹, LUO Jinling¹

1. School of Electrical and Information Technology, Yunnan Minzu University, Kunming 650500, China;
2. Program for Innovative Research Team (in Science and Technology) in University of Yunnan Province, Kunming 650500, China

关键词:

轻度认知障碍; rs-fMRI; 迁移学习; 阿尔茨海默症; MobileNet; 深度学习; 机器学习; 兴趣区

Keywords:

EMCI; rs-fMRI; transfer learning; Alzheimer’s disease; MobileNet; deep learning; machine learning; region of interest

分类号:

TP391.41

DOI:

10.11992/tis.202007041

摘要:

早期轻度认知障碍(early mild cognitive impairment, EMCI)是阿尔兹海默症(Alzheimer’s disease, AD)发生前的必经阶段，EMCI的准确诊断对AD早期预防和治疗都具有积极作用。通过静息态功能性磁共振成像(rest-state functional magnetic resonance imaging，rs-fMRI)来诊断EMCI的深度学习方法性能主要依赖如何提取特征值，而传统提取方法存在特征信息易损失和训练网络时间过长等问题。针对该问题，本文采用迁移学习的特征提取方法来对EMCI分类，首先从rs-fMRI中提取兴趣区(region of interest, ROI)时间序列，以此完成源数据的降维，然后利用MobileNet迁移网络从ROI中提取瓶颈特征，最后将该特征输入到设计的分类顶层网络以得到最后分类结果。在实验中，采用阿尔茨海默病神经影像学数据库(Alzheimer’s disease neuroimaging initiative，ADNI)的数据进行测试，实验结果表明，本文的方法比传统方法的分类精度提高了约10%，而分类时间大约只有传统方法的25%。

Abstract:

Early mild cognitive impairment (EMCI) is a necessary stage before Alzheimer’s disease (AD). Thus, accurate diagnosis of EMCI will help the early prevention and treatment of AD. The performance of the deep learning method for the diagnosis of EMCI through rest-state functional magnetic resonance imaging (rs-fMRI) mainly depends on how feature values are extracted. However, traditional extraction methods have problems, such as loss of feature information and long training network time. To address these problems, this paper proposes a transfer learning-based feature extraction method for EMCI classification. First, the region of interest (ROI) time series is extracted from rs-fMRI to complete the dimensional reduction of source data. Second, bottleneck features are extracted from the ROI by using a transfer learning MobileNet. Lastly, a final classification result is obtained from those entered features into a designed top-layer net. The data of the Alzheimer’s Disease Neuroimaging Initiative are tested through experiments. Experimental results show that compared with traditional methods, the classification accuracy of the proposed method is enhanced by around 10%, and 75% or so of classification time is saved.