基于卷积神经网络的白细胞分类方法研究  

AResearch on White Blood Cell Classification Method Based on Convolutional Neural Network

作　　者：宋小锋 赵宇 SONG Xiao-feng;ZHAO Yu(College of Electronic Information and Electrical Engineering,Tianshui Normal University,Tianshui 741001,China)

机构地区：[1]天水师范学院电子信息与电气工程学院,甘肃天水741001

出　　处：《电脑与信息技术》2024年第5期17-20,70,共5页Computer and Information Technology

基　　金：天水师范学院2021年创新基金资助项目(项目编号:CXJ2021-26);天水师范学院大学生创新创业训练计划项目(项目编号:CX20220047)。

摘　　要：白细胞进行分类的传统方法在特征提取中存在耗时、准确率不高的问题。基于此,提出利用深度学习方法实现特征自动提取,并使用开源的BCCD数据集进行测试。数据集共有12 436张白细胞图像,分别使用经典的卷积神经网络AlexNet、VGG11、GoogLeNet、ResNet18、ResNet34、DenseNet121、EfficientNetB0模型对图片进行训练,通过混淆矩阵定量评价模型训练结果。结果表明,EfficientNetB0网络模型总体上优于其他网络,准确率达到96.02%,且在测试集上耗时仅为8.3 s。为了提高模型的可解释性,可视化了类热力图,同时证实了实验结果。The traditional method of white blood cell classification has the problem of time-consuming and low accuracy in feature extraction.It is proposed to use deep learning methods to achieve automatic feature extraction and use open source BCCD data sets for testing.There are 12436 white blood cell,images in the data set.The classic convolutional neural networks AlexNet,VGG11,GoogLeNet,ResNet18,ResNet34,DenseNet121,and EfficientNetB0 models are used to train the images,and the model training results are quantitatively evaluated by the confusion matrix.The results show that the EfficientNetB0 network model is generally superior to other networks,with an accuracy rate of 96.02%and a time of only 8.3 s on the test set.In order to improve the interpretability of the model,the heat map is visualized and the experimental results are also confirmed.

关 键 词：白细胞 深度学习 图像分类 可视化 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]