东北大学学报(自然科学版) ›› 2024, Vol. 45 ›› Issue (12): 1673-1679.DOI: 10.12068/j.issn.1005-3026.2024.12.001

• 信息与控制 •    

融合多尺度注意力机制的冠状病毒肺炎CT诊断方法

单鹏, 张林(), 肖洪明, 赵玉良   

  1. 东北大学秦皇岛分校 控制工程学院,河北 秦皇岛 066004
  • 收稿日期:2023-07-03 出版日期:2024-12-10 发布日期:2025-03-18
  • 通讯作者: 张林
  • 作者简介:单 鹏(1985-),男,河南平舆人,东北大学副教授.
  • 基金资助:
    中央高校基本科研业务费专项资金资助项目(N2023021)

CT Diagnosis Method for Coronavirus Pneumonia with Integrated Multi-scale Attention Mechanism

Peng SHAN, Lin ZHANG(), Hong-ming XIAO, Yu-liang ZHAO   

  1. School of Control Engineering,Northeastern University at Qinhuangdao,Qinhuangdao 066004,China.
  • Received:2023-07-03 Online:2024-12-10 Published:2025-03-18
  • Contact: Lin ZHANG

摘要:

人工智能诊断是肺部感染的重要辅助诊断方法之一.然而,现有的方法大多基于深度学习,具有模型稳定性不足、复杂度高、准确率低的问题.提出融合多尺度注意力机制的浅层模型,实现了准确率高并且结构简单的新型冠状病毒肺炎CT诊断.首先,将收集到的两个新型冠状病毒肺炎数据集融合成一个数据集,解决了因数据集不足导致训练的模型不稳定.其次,通过在浅层网络ResNet18后3层中引入多尺度注意力,弥补了模型特征提取能力的不足.最后,搭建了一个具有3层全连接层的分类器,改进模型的分类能力,进而提高了肺部CT的分类准确率.结果表明,所提模型准确率达到95.41%,性能超过ResNet50,ResNet101,VGG16,DenseNet169等网络,并且模型参数数量仅有12.24×106,比ResNet50和VGG16等网络低50%左右.

关键词: 肺炎, 深度学习, 多尺度注意力, CT, 分类器

Abstract:

Artificial intelligence (AI)‑based diagnosis has become an important auxiliary method for detecting lung infections. However, most existing approaches rely on deep learning, which are often plagued by issues such as insufficient model stability, high complexity, and low accuracy. This paper presents a shallow model which incorporates a multi‑scale attention mechanism to achieve both high accuracy and a simple structure for diagnosing COVID‑19 from CT scans. Firstly, two datasets of COVID‑19 CT images are combined into a single dataset to address the issue of model instability caused by insufficient data. Secondly, by introducing multi‑scale attention(MA) in the final three layers of the shallow ResNet18 network, the model’s feature extraction capability is enhanced. Finally, classifier with three fully connected layers (CTFCL) is constructed to improve the classification performance of the model, thereby increasing the accuracy of lung CT classification. Experimental results show that the proposed model achieves an accuracy of 95.41%, outperforming networks such as ResNet50, ResNet101, VGG16, and DenseNet169. Furthermore, the model has only 12.24×106 parameters, which is approximately 50% fewer than networks like ResNet50 and VGG16.

Key words: pneumonia, deep learning, multi-scale attention, CT, classifier

中图分类号: