基于Transformer模型的自闭症功能磁共振图像分类

PDF下载HTML

Title:

Classification of functional magnetic resonance images for autism based on Transformer model

作者:

潘登¹, 毕晓君^2,3

1. 哈尔滨工程大学 信息与通信工程学院, 黑龙江 哈尔滨 150001;
2. 民族语言智能分析与安全治理教育部重点实验室, 北京 100081;
3. 中央民族大学 信息工程学院, 北京 100081

Author(s):

PAN Deng¹, BI Xiaojun^2,3

1. College of Information and Communication Engineering, Harbin Engineering University, Harbin 150001, China;
2. Key Laboratory of Ethnic Language Intelligent Analysis and Security Governance of MOE, Beijing 100081, China;
3. Department of Informati

关键词:

深度学习; Transformer; 注意力机制; 自闭症; 功能磁共振成像; 图像分类; 特征提取; 功能连接

Keywords:

deep learning; Transformer; attention mechanism; autism; functional magnetic resonance imaging; image classification; feature extraction; functional connectivity

分类号:

TP391

DOI:

10.11992/tis.202402025

摘要:

目前自闭症功能磁共振(functional magnetic resonance imaging, fMRI)图像分类模型在跨多个机构的数据集下分类精度较低，难以应用到自闭症的诊断工作中。为此，本文提出了一种基于Transformer的自闭症分类模型(autism spectrum disorder classification model based on Transformer, TransASD)。首先采用脑图谱模板提取fMRI数据中的时间序列输入Transformer模型，并引入一种重叠窗口注意力机制，能够更好地捕捉异构数据的局部与全局特征。其次，提出了一个跨窗口正则化方法作为额外的损失项，使模型可以更加准确地聚焦于重要的特征。本文使用该模型在公开的自闭症数据集ABIDE上进行实验，在10折交叉验证法下得到了71.44%的准确率，该模型对比其他先进算法模型取得了更好的分类效果。

Abstract:

Current classification models of functional magnetic resonance (fMRI) images for autism struggle with low classification accuracy across datasets from multiple institutions. Thus, they have difficulty assisting in the diagnosis of autism. This study proposes a Transformer-based autism classification model named TransASD to address this issue. This model utilizes brain mapping templates to extract time series from fMRI data and incorporates an overlapping window attention mechanism to better capture local and global features of heterogeneous data. A cross-window regularization method is also proposed as an additional loss term, which allows the model to focus more accurately on important features. In this study, we use the model to conduct experiments on the publicly available autism dataset ABIDE, under the ten-fold cross-validation method, the accuracy rate is 71.44%. Experimental results show that the model achieves state-of-the-art performance compared with other advanced algorithmic models.