基于双重注意力机制的间质性肺病高分辨率CT图像分类方法  

High resolution CT image classification method for interstitial lung disease based on dual attention mecha

作　　者：赵琪玉 张俊华[1] 张剑青[2] 徐铭蔚 Zhao Qiyu;Zhang Junhua;Zhang Jianqing;Xu Mingwei(School of Information,Yunnan University,Kunming 650500,China;The First Affiliated Hospital of Kunming Medical University,Kunming 650032,China)

机构地区：[1]云南大学信息学院,昆明650500 [2]昆明医科大学第一附属医院,昆明650032

出　　处：《国外电子测量技术》2024年第6期1-11,共11页Foreign Electronic Measurement Technology

基　　金：国家自然科学基金(62063034,61841112)项目资助。

摘　　要：为了更精确地分类间质性疾病,提出了一种基于深度学习的分类网络,首先将多头自注意力机制模DenseNet-121结合,使得模型能够同时关注多个重点区域。然后采用卷积注意力模块实现更高效的特征提取,提升网络的空间感知能力,从而增强分类性能。最后,添加改进的空间金字塔池化层将不同尺度的特征图拼接起来以捕获更丰富的空间信息。此外针对高分辨率C图像数据集类别不均衡问题,引入FocalLoss损失函数,使得模型在训练时更专注于难分类的样本,从而进一步增强模型的分类能力。所提方法在未经训练的数据集上进行测试,达到了88.28%的准确率。相较于原始DenseNet-121在准确率、召回率、精确率、F1分数和Kappa系数提高了4.65%、5.08%、5.82%、5.45%和6.38%。实验结果表明,该方法具有特征提取能力强和分类准确率高的特点。In order to classify interstitial diseases more accurately,this paper proposes a classification network based on deep learning Xi,which first combines the multi-head self-attention mechanism module with DenseNet-121,so that the model can focus on multiple key regions at the same time.Then,the convolutional attention module is used to achieve more efficient feature extraction and spatial perception capabilities,so as to improve the classification ability of the network.Finally,an improved spatial pyramid pooling layer is added to stitch together the feature maps of different scales to capture richer spatial information.In addition,aiming at the problem of category imbalance of high-resolution CT image datasets,the Focal Loss function is introduced,which makes the model focus more on the difficult samples during training,so as to further improve the classification ability of the model.The proposed method is tested on the untrained dataset in this paper,and the accuracy rate reaches 88.28%.Compared with the original DenseNet-121,the accuracy,recall,precision,F1 score and Kappa coefficient are increased by 4.65%,5.08%,5.82%,5.45%and 6.38%.Experimental results show that the proposed method has the characteristics of strong feature extraction ability and high classification accuracy.

关 键 词：间质性肺病 深度学习 注意力机制 DenseNet-121 高分辨率CT图像 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]