基于多尺度深度卷积神经网络的骨髓白细胞识别研究

Research of bone marrow white blood cell recognition based on multi-scale deep convolutional neural network

针对骨髓白细胞人工提取特征复杂,识别困难等问题,提出一种多尺度滤波深度卷积神经网络(MS-DCNN)模型。首先,该模型将传统的卷积神经网络模型的滤波器尺寸缩小,以减少模型的总体参数以提升网络模型训练的效率;其次,通过增加滤波器的个数和网络深度来提取骨髓血细胞更丰富的特征;最后通过在Sysmex(希森美康)公开数据集上对6类骨髓白细胞进行实验,并与其他主流分类方法进行对比。结果表明,该文提出的MS-DCNN模型准确率达到了98.9%,高于其他主流方法,其有效性得到了验证。

A multi⁃scale filtering depth convolutional neural network(MS⁃DCNN)model is proposed to improve the problems of complex features and difficult recognition of manual extraction of the bone marrow white blood cells.In this MS⁃DCNN model,the filter size of the traditional convolutional neural network model is shrinked to reduce the overall parameters of it,so as to improve the efficiency of network model training.The more abundant features of bone marrow blood cells can be extracted by increasing the number of filters and the network depth.The experiment of six types of bone marrow white blood cells are conducted on the Sysmex open datasets,and compared with other mainstream classification methods.The results show that the accuracy of the proposed MS⁃DCNN model is up to 98.9%,higher than other mainstream methods,and its effectiveness has been verified.