| 吕浩音,周泽华,孙楠,李巧慧,张占平,刘变叶.融合高效通道注意力机制改进轻量化网络模型基于肺部X线片分类正常肺与不同病原体所致肺炎[J].中国医学影像技术,2026,42(6):951~956 |
| 融合高效通道注意力机制改进轻量化网络模型基于肺部X线片分类正常肺与不同病原体所致肺炎 |
| Fusing efficient channel attention to improved lightweight network model for classifying normal lung and pneumonia caused by different etiologies based on chest X-ray films |
| 投稿时间：2025-10-23 修订日期：2026-02-25 |
| DOI：10.13929/j.issn.1003-3289.2026.06.033 |
| 中文关键词: 肺炎 X线 人工智能 |
| 英文关键词:pneumonia X-rays artificial intelligence |
| 基金项目:甘肃省科技计划(23JRRM0746)、庆阳市科技重大专项(2025JY1003)。 |
|
| 摘要点击次数: 149 |
| 全文下载次数: 58 |
| 中文摘要: |
| 目的 提出融合高效通道注意力机制(ECA)的改进轻量化网络(MobileNetV3-ECA)模型,观察其基于胸部X线片分类正常肺与不同病原体所致肺炎的价值。方法 基于采自肺部无异常、细菌性肺炎、新型冠状病毒感染(COVID-19)及其他病毒性肺炎患者的共4 360幅胸部X线片构建数据集,并按9∶1比例划分训练集(n=3 925)与验证集(n=435)。于MobileNetV3主干结构中嵌入ECA模块,构建MobileNetV3-ECA轻量化网络模型;与MobileNetV2、ResNet18、RestNet34、VGG11及GoogleNet模型对比,观察该模型分类正常肺与不同病原体所致肺炎的效能。结果 MobileNetV3-ECA模型大小4.27 MB、推理时间16.75 ms,显著低于其他模型;其分类COVID-19和其他病毒性肺炎的精确率最高,分别为81.31%及95.28%;分类正常肺的精确率为79.81%,不及MobileNetV2、VGG11及GoogleNet模型(84.72%~100%);分类细菌性肺炎精确率为86.36%,仅次于ResNet18模型(93.02%)。结论 利用MobileNetV3-ECA模型可在基于胸部X线片准确分类正常肺与不同病原体所致肺炎的同时显著提升计算效率。 |
| 英文摘要: |
| Objective To propose a fusing efficient channel attention (ECA) to improved lightweight network (MobileNetV3-ECA) model, and to evaluate its value for classifying normal lung and pneumonia caused by different etiologies based on chest X-ray films. Methods A dataset comprising 4 360 chest X-ray films acquired from patients with normal lungs, bacterial pneumonia, coronavirus disease 2019 (COVID-19) and other viral pneumonia was constructed. Training set (n=3 925) and validation set (n=435) were divided at a ratio of 9∶1. ECA module was embedded into the backbone architecture of MobileNetV3 to construct the lightweight MobileNetV3-ECA model, and the performance of this model for classifying normal lung and pneumonia caused by different etiologies were evaluated and compared with that of MobileNetV2, ResNet18, ResNet34, VGG11 and GoogLeNet models. Results MobileNetV3-ECA model had a size of 4.27 MB and an inference time of 16.75 ms, both were markedly lower than those of the other models. MobileNetV3-ECA model achieved the highest precision for classifying COVID-19 (81.31%) and other viral pneumonia (95.28%), respectively. The precision of MobileNetV3-ECA model for classifying normal lung was 79.81%, lower than that of MobileNetV2, VGG11 and GoogLeNet models (84.72%—100%), for identifying bacterial pneumonia was 86.36%, second only to that of ResNet18 model (93.02%). Conclusion MobileNetV3-ECA model could substantially enhance computational efficiency while ensure the accuracy of classifying normal lung and pneumonia caused by different etiologies based on chest X-ray films. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
