《计算机应用》唯一官方网站 ›› 2023, Vol. 43 ›› Issue (S1): 26-32.DOI: 10.11772/j.issn.1001-9081.2022081240

• 人工智能 • 上一篇    下一篇

利用3D-RepVGG进行阿尔兹海默症诊断

胡众义1,2(), 张夏彬1,2   

  1. 1.温州大学 计算机与人工智能学院,浙江 温州 325035
    2.温州市智能影像处理与分析重点实验室(温州大学),浙江 温州 325035
  • 收稿日期:2022-08-22 修回日期:2022-10-22 接受日期:2022-11-03 发布日期:2023-07-04 出版日期:2023-06-30
  • 通讯作者: 胡众义
  • 作者简介:胡众义(1977—),男,浙江温州人,教授,博士,CCF高级会员,主要研究方向:机器视觉、机器学习、智能信息处理 hujunyi@163.com
    张夏彬(1998—),男,浙江绍兴人,硕士研究生,主要研究方向:机器学习、计算机辅助诊断。
  • 基金资助:
    浙江省自然科学基金资助项目(LD21F020001);国家自然科学基金重点支持项目(U1809209);温州市科技计划重大科技创新攻关项目(ZY2019020)

Alzheimer disease diagnosis by three-dimension re-parameterization visual geometry group

Zhongyi HU1,2(), Xiabin ZHANG1,2   

  1. 1.College of Computer Science and Artificial Intelligence,Wenzhou University,Wenzhou Zhejiang 325035,China
    2.Key Laboratory of Intelligence Image Processing and Analysis (Wenzhou University),Wenzhou Zhejiang 325035,China
  • Received:2022-08-22 Revised:2022-10-22 Accepted:2022-11-03 Online:2023-07-04 Published:2023-06-30
  • Contact: Zhongyi HU

摘要:

阿尔兹海默症(AD)临床症状为失忆、失语与丧失行动能力等。AD暂无有效治疗方法,但早期干预已证明有效,因此,AD早期诊断至关重要。针对该问题,基于RepVGG网络架构中的结构重参数化技术,将训练阶段的多分支卷积网络等效转换为预测阶段的单分支卷积网络,获得多分支卷积网络性能高与单分支卷积网络速度快等优点;同时,利用3D卷积引入空间连续信息;最终,成功地将RepVGG网络架构与3D卷积融合,提出3D-RepVGG网络,以实现对AD、轻度认知障碍(MCI)和正常对照组(NC)的诊断。实验数据来自于公开数据库ADNI,原始的磁共振图像(MRI)使用SPM12进行预处理。预处理后数据输入3D-RepVGG进行AD/NC、MCI/NC、AD/MCI、AD/MCI/NC四种分类任务,分别获得了90.38%、85.90%、70.51%、62.50%的准确率。实验结果表明,3D-RepVGG在AD早期诊断任务上能获得较好的诊断结果。

关键词: 阿尔兹海默症, 计算机辅助诊断, 三维卷积神经网络, 多分支卷积神经网络, 图像分类

Abstract:

Alzheimer Disease (AD) is characterized by memory loss, aphasia and loss of mobility. There is no effective treatment for AD, but early intervention has proven to be effective. Therefore, early diagnosis of AD is crucial. To address this problem, based on the structural reparameterization technique in the RepVGG (Re-parameterization Visual Geometry Group) network architecture, the multi-branch convolutional network in the training phase was equivalently converted into a single-branch convolutional network in the prediction phase, and the advantages of high performance of multi-branch convolutional network and fast speed of single-branch convolutional network were obtained. Meanwhile, 3D convolution was used to introduce spatial continuous information. Finally, the RepVGG network architecture was successfully fused with 3D convolution to propose 3D-RepVGG network for the diagnosis of Alzheimer, Mild Cognitive Impairment (MCI) and Normal Control (NC). The experimental data were obtained from the publicly available database ADNI. Raw Magnetic Resonance Imaging (MRI) images were preprocessed using SPM12. The preprocessed data were put into the 3D-RepVGG network for four classification tasks, AD/NC, MCI/NC, AD/MCI and AD/MCI/NC. Finally, 3D-RepVGG obtained 90.38%, 85.90%, 70.51%, and 62.50% accuracy on the four classification tasks, respectively. The experimental results show that 3D-RepVGG obtains great diagnostic results on the AD early diagnosis task.

Key words: Alzheimer Disease (AD), computer aided diagnosis, 3D Convolution Neural Network (CNN), multi-branch CNN, image classification

中图分类号: