摘要
针对当前卷积神经网络在医学CT图像肺部结节分类中存在图像特征提取不全面,导致分类准确度低且检测时间长的问题,提出了一种基于深度网络特征融合的分类检测网络(efficient selective convolutional network,ESC-Net),网络是以EfficientNet-V1为基础框架,在MBConv(mobile inverted residual bottleneck convolution)结构中引入轻量级注意力机制,同时,为降低网络的参数量和FLOPs,删去3层MBConv结构,进一步增强了特征提取和分类能力,适合于实际应用场景中快速、精准地诊断恶性结节。结果表明,在LIDC-IDRI数据集上,方法实现了对肺结节良恶性的精确分类,分类准确率和AUC值分别达到了94.6%与98.3%,优于大部分主流的分类方法。
In view of the problems of incomplete feature extraction in the classification of medical CT images by convolutional neural network,resulting in low classification accuracy and long detection time,An efficient selective convolutional network(ESC-Net)based on deep network feature fusion is proposed in this paper.Based on the EfficientNet-V1 framework,the network introduces a lightweight attention mechanism in its MBConv structure.In order to reduce the number of parameters and FLOPs of the network,the three-layer MBConv structure is deleted to further enhance the feature extraction and classification capability,which is suitable for the rapid and accurate diagnosis of malignant nodules in practical application scenarios.The results showed that,on the LIDC-IDRI dataset,the proposed method could accurately classify the benign and malignant pulmonary nodules,with the classification accuracy and AUC value reaching 94.6%and 98.3%,respectively,which was superior to most mainstream classification methods.
作者
禹文明
刘伟
张其超
Yu Wenming;Liu Wei;Zhang Qichao(School of Automation,Nanjing University of Information Science&.Technology,Nanjing 210044,China;School of Internet of Things,Wuxi University,Wuxi 214105,China)
出处
《国外电子测量技术》
北大核心
2023年第7期181-187,共7页
Foreign Electronic Measurement Technology
基金
江苏高校哲学社会科学研究项目(2021SJA2307)
上海市分子影像学重点实验室开放课题(KT-2022-17)项目资助。
