基于改进ResNet的多标签胸部X光图像分类  

Multi-label chest X-ray image classification based on improved ResNet

作　　者：方燕燕 陈辉[1] FANG Yanyan;CHEN Hui(School of Computer Science and Engineering,Anhui University of Science and Tenchnology,Huainan 232001,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《山东理工大学学报（自然科学版）》2025年第2期1-9,共9页Journal of Shandong University of Technology:Natural Science Edition

基　　金：国家自然科学基金项目(61170060)。

摘　　要：针对目前胸部X光图像分类方法存在X光特异性特征表达不充分、高频图像特征提取效果差、疾病样本数量不平衡等问题,提出一种基于改进ResNet的多标签胸部X光图像分类方法(multi-label chest X-ray image classification based on improved ResNet, MLC-ResNet)。首先,设计一个多尺度特征提取和融合模块,以获取更丰富的特征信息;其次,将八度卷积替换为残差结构中的普通卷积,解决X光特异性特征表达不充分问题;再次,为改善高频特征提取效果,在ResNet中引入改进后的多层感知器(multilayer perception, MLP),更好地揭示图像的细节和整体结构,增加分类性能;最后,使用加权交叉熵损失函数增加样本数较少的类别权重,改善样本分布不平衡问题。在ChestX-Ray14和CheXpert数据集上进行实验测试,其平均AUC分别是0.858 7和0.844 7,相较于ResNet分类算法分别提高了4.47%和3.20%。通过与现有方法的对比实验,进一步证明MLC-ResNet模型具有更好的性能。In light of the limitations in current chest X-ray image classification methods,such as inadequate representation of specific X-ray image features,suboptimal feature extraction for high-frequency images,and imbalanced distribution of disease samples,a novel approach termed multi-label chest X-ray image classification based on improved ResNet(MLC-ResNet)is proposed.First,a multi-scale feature extraction and fusion module is developed to capture richer feature information.Second,to enhance the expression of X-ray specific features,octave convolution is replaced by common convolution in the residual structure.Additionally,an enhanced multilayer perception(MLP)is integrated into the ResNet framework to optimize high-frequency feature extraction,and better reveal visualization of image details and overall structure,which collectively bolsters the classification performance.Last,the weighted crossentropy loss function is employed to address the sample imbalance issue by assigning greater weights to classes with fewer samples.Experiments conducted on the ChestX-Ray14 and CheXpert datasets yielded average AUC scores of 0.8587 and 0.8447,representing improvements of 4.47%and 3.20%over traditional ResNet classification methods,respectively.Comparative analyses with existing approaches further demonstrate the superior performance of the MLC-ResNet model.

关 键 词：多层感知器 ResNet 加权交叉熵损失函数 胸部X光图像 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]