Automatic White Blood Cell Classification Using Deep Learning Network

Changes in white blood cell, a.k.a. leukocyte, subtypes count are often indicator of medical conditions. Counting each of white blood cell subtypes is thus essential in blood tests. To do this, it is necessary to classify white blood cell images into different subtypes, namely, Eosinophils, Lymphocytes, Monocytes, Neutrophilic, etc. Traditional image processing techniques have been utilized to classify white blood cell images. However, in general, accuracy and stability of these techniques are sensitive to variations in image size, staining technique, and image quality. Recently, application of deep learning networks in white blood cell classification has drawn a lot of interest. In this work, we present two deep learning networks based on VGG16 and SqueezeNet convolutional neural networks (CNNs) that can automatically classify white blood cell images into different subtypes. The two networks were trained on a public dataset of leukocytes named Blood Cell Count and Detection (BCCD) and a self-built dataset named the Extended BCCD dataset. Although having relatively light-size, 16,005,884 and 35,858 parameters for VGG16 and SqueezeNet respectively, both models achieved ~ 97% accuracy in the BCCD dataset. In the extended BCCD dataset, the VGG16-based model achieved 97% accuracy, and the SqueezeNet-based model achieved 95% accuracy. These two models are thus potential candidates for deployment on computers and IoT devices with limited computational power for white blood cell classification task.