结合自注意力与卷积的胸部X光片疾病分类研究
Research on Combining Self-Attention and Convolution for Chest X-Ray Disease Classification
关欣 1耿晶晶 1李锵1
作者信息
- 1. 天津大学微电子学院,天津 300072
- 折叠
摘要
胸部X光片可用于诊断多种胸部疾病.由于胸部疾病特征复杂多样,现有的胸部X光片疾病分类算法难以学习胸部疾病复杂的鉴别表征以及未关注不同疾病之间的相关性信息.针对以上问题,提出一种结合自注意力与卷积的疾病分类算法,该算法采用全维度动态卷积替换残差网络的标准卷积,从而提高网络对多尺度信息的特征提取能力.此外,在卷积神经网络中引入自注意力模块,可以提供捕获多种疾病之间相关性的全局感受野.最后,提出高效的双路注意力,使神经网络更加关注病灶区域、自动捕捉病变位置变化.在ChestX-ray14数据集上,对所提模型进行评估,实验结果表明:所提算法对14种胸部疾病的平均受试者工作特性曲线下的面积(AUC)达到0.839,检测结果与目前其他7种先进算法相比在准确率和效率上有所提升.
Abstract
Chest X-rays are used to diagnose a wide range of chest conditions.However,due to the complicated and diverse features of thoracic diseases,existing disease classification algorithms for chest radiographs have difficulty in learning the complex discriminating features of thoracic diseases and do not fully consider correlation information between different diseases.This study proposes a disease classification algorithm that combines self-attention and convolution to address these problems.This study employs omni-dimensional dynamic convolution to replace the standard convolution of the residual network to enhance the feature extraction capabilities of the network for multi-scale information.In addition,a self-attention module is introduced into the convolutional neural network to provide global receptive fields that capture correlations between multiple diseases.Finally,an efficient double path attention is proposed that allows the network to give greater attention to the focal area and automatic capturing of changes in lesion locations.The proposed model is evaluated on the ChestX-ray14 dataset.Experimental results show that the accuracy of the algorithm and the efficiency of diagnosis for the classification of 14 chest diseases is improved over those of the seven current state-of-the-art algorithms,with an average area under receiver operating characteristic curve(AUC)value of 0.839.
关键词
胸部X光片/全维度动态卷积/自注意力/双路注意力/疾病分类Key words
chest X-ray/omni-dimensional dynamic convolution/self-attention/double path attention/disease classification引用本文复制引用
基金项目
国家自然科学基金(61471263)
国家自然科学基金(61872267)
国家自然科学基金(U21B2024)
天津市自然科学基金(16JCZDJC31100)
天津市研究生科研创新项目(2021YJSS023)
天津大学创新基金会项目(2021XZC-0024)
出版年
2024
NETL
NSTL
万方数据