摘要
白细胞图像的自动分类有助于提高临床诊疗效率,但仍需进一步改进方法以提高分类正确率。探索用卷积神经网络(CNN)进行外周血白细胞图像的自动分类识别。在深度学习框架Caffe上,以Alex Net和Le Net为网络原型构建CNN训练平台;用Cella Vision DM96采集外周血涂片中的5类白细胞图像,经人工鉴定后按训练∶校验∶测试=7∶2∶1的比例,随机分配图像构建原始数据集,再通过平移、旋转及镜像构建扩充数据集;训练时采用随机梯度下降算法优化模型权值,以分类准确率>95%为目标评估训练结果及优化调整网络结构。结果发现,Alex Net的训练误差无法收敛,陷入局部极小,Le Net则达到预期目标。随后对Le Net网络进行删减优化,获得一轻量高效的新结构——CCNet,其在模型大小、训练用时和分类用时上分别仅为Le Net的1/1000、1/3和1/30。两者对979张5类细胞图像的最佳分类准确率分别达到99.69%和99.18%,高于目前同类研究报道。结果表明,CNN可用于5类白细胞图像的"端对端"分类识别,特别是CCNet模型兼具准确与效率优势。
The automatic classification of white blood cell（ WBC） image is essential because it helps to enhance the efficiency of clinical diagnosis and treatment. However,the classification accuracy is still need to be boost for adapting to practical applications. In this paper,we proposed an automatic classification method based on the convolution neural network（ CNN）. We tentatively fed our training dataset into Alex Net and Le Net using a widely-used deep learning platform Caffe. Five classes of WBCs images collected by a Cella Vision DM96 in peripheral blood smears were adopted as the training dataset. These manual labeled images were apportioned into three groups（ training,validation and testing） randomly to construct the original dataset according to the proportion of7∶ 2∶ 1. With the augmentation methods,such as rotation and mirror,we expanded the original dataset. Stochastic gradient descent algorithm was adopted as the optimizing method for training CNNs. The experimental results demonstrated that the network structure of Alex Net was unsuitable to achieve the ideal classification accuracy which more than 95%. While the network structure of Le Net had achieved the expected target. However,the more massive and more time-consuming of Le Net suggested us to further optimize the connection of layers to derive a new network with lightweight structure,named as CCNet. The model size,time for training,and time for evaluation of CCNet were only 1/1000,1/3,and 1/30 compared with Le Net,respectively. The best classification accuracy of CCNet and Le Net for five-classification of WBCs was 99. 69% and 99. 18% with 979 WBC images,higher than those of the previous reports. It demonstrated that CNNs especially CCNet had clear advantages thanprevious works both in classification accuracy and speed.
出处
《中国生物医学工程学报》
CAS
CSCD
北大核心
2018年第1期17-24,共8页
Chinese Journal of Biomedical Engineering
基金
福建省科技重点项目(2012Y0058)
关键词
深度学习
卷积神经网络
白细胞形态
分类识别
deep learning
convolution neural network
leukocyte morphology
classification
