一种白血细胞图像训练集扩充方法  

A method for white blood cell image training set augmentation

作　　者：臧宇 苏洋 ZANG Yu;SU Yang(Department of Hematology,Huizhou First Hospital,Huizhou 516000,China;School of Computer and Information,Dongguan City College,Dongguan 523419,China)

机构地区：[1]惠州市第一人民医院血液内科,广东惠州516000 [2]东莞城市学院计算机与信息学院,广东东莞523419

出　　处：《中国医学物理学杂志》2023年第3期342-349,共8页Chinese Journal of Medical Physics

基　　金：广东省教育厅重点领域专项(2021ZDZX1029);东莞市社会科技发展(重点)项目(2020507151144)。

摘　　要：针对因训练集较小导致的白血细胞图像识别精度低以及传统的扩充训练集方法需要人工介入的问题,提出一种白血细胞图像训练集扩充方法,将图像旋转任意角度后,提取因旋转产生的黑色区域边缘,然后对黑色区域进行填充,并弱化边缘特征,得到扩充训练集。实验结果表明,使用本文方法扩充训练集对ResNet50、MobileNet与ShuffleNet 3种模型进行训练后,对比原始数据集,模型的识别精度分别提高220.18%、140.84%与88.99%,且不需要人工介入。Aiming at the low recognition accuracy of white blood cell images caused by the small training set and the need for manual intervention in the traditional method of training set augmentation,a novel method is proposed for the augmentation of the training set of white blood cell images.The edges of the black area caused by the image rotation by an arbitrary angle are extracted,and the training set augmentation is realized through filling the black area and weakening the edge features.The experimental results show that after using the training set augmented by the proposed method to train ResNet50,MobileNet and ShuffleNet,comparing with the original data set,the recognition accuracies of these models are improved by an average of 220.18%,140.84%,88.99%,and no manual intervention is required.

关 键 词：白血细胞识别 机器学习 训练集扩充 

分 类 号：R318[医药卫生—生物医学工程] TP3[医药卫生—基础医学]