文章摘要

文章信息/Info

Title:: Diagnosis of Alzheimer′s disease based on 3D convolutional neural network-regions of interest

作者:: 曾安1; 2; 邹超1; 潘丹3△; 1.广东工业大学计算机学院, 广州 510006；2.广东大数据分析与处理重点实验室，广州 510006；3.广州建设职业技术学院现代教育技术中心，广州 510440

Author(s):: ZENG An1; 2; ZOU Chao1; PAN Dan3; 1. School of Computers, Guangdong University of Technology, Guangzhou 510006, China;2. Guangdong Key Laboratory of Big Data Analysis and Processing, Guangzhou 510006;3. Modern Education Techonology Center，Guangdong Construction Polytechnic, Guangzhou

Keywords:: Alzheimer′s disease; Classification; Convolutional neural network; Region of interest; Cross validation

摘要:: 磁共振(magnetic resonance imaging,MRI)图像的预测分类对早期阿尔茨海默症(Alzheimer′s disease,AD)的诊断非常重要。轻度认知障碍(mild cognitive impairment,MCI)作为AD的一种早期阶段，在诊断时存在大脑脑区萎缩区域不明确，诊断准确率偏低等问题。本研究提出一种基于感兴趣区域(regions of interest,ROI)的3D卷积神经网络(convolutional neural network, CNN)模型来解决AD分类准确率偏低等问题，进而实现对AD的计算机辅助诊断。实验数据均来自ADNI数据库，实验结果表明，基于ROI的3D CNN的AD辅助诊断模型在分类AD vs 正常对照(normal control,NC)、MCI转化AD(MCI converted to AD,MCIc) vs NC和MCI未被转化AD(MCI not converted to AD,MCInc) vs MCIc的5折交叉验证平均准确率分别为85.2%、83.9%、68.5%。相比于传统的主成分分析+支持向量机方法和单纯的切片集成方法，本研究方法在AD辅助诊断中取得了更好的分类效果和泛化能力，还可为其他脑疾病诊断提供新思路。

Abstract:: The predictive classification of magnetic resonance (MRI) images is very important for the diagnosis of Alzheimer′s disease (AD). Mild cognitive impairment (MCI) as an early stage of AD with the problem of unclear area of brain atrophy and low accuracy at diagnosis. We proposed a 3D convolutional neural network (CNN) model based on regions of interest (ROI) to solve the low accuracy of AD prediction, then to achieve computer-aided diagnosis of AD. Experimental data was obtained from the ADNI database,the experimental results showed that the 5-fold cross-validation average accuracy of AD vs NC, MCIc(MCI converted to AD)vs NC and MCInc(MCI not converted to AD) vs MCIc in the diagnosis model of 3D CNN-ROIs respectively reached 85.2%, 83.9% and 68.5%. Compared with the PCA+SVM method and slices ensemble method, the results achieve better classification and generalization ability in the diagnosis of AD, and may provide new ideas for the diagnosis of other brain diseases.