Automatic Classification of White Blood Cells Using a Semi-Supervised Convolutional Neural Network

The correct classification of white blood cell subtypes is critical in the diagnosis of blood disease. However, the performance of classical computer vision-based classification methods is heavily dependent on the features that should be carefully designed by trial and error. The machine learning-based classifier outperforms the traditional classifiers but suffers from sample labeling, which is labor intensive and time consuming. This paper presents a semi-supervised convolutional neural network that can maintain a similarly high accuracy of classification as deep learning approaches with only 10% labeled data or less. A Visual Geometry Group (VGG) network model was pre-trained with a small amount of labeled data and then used to predict unlabeled data. After implementing entropy filtering and confidence filtering processes, high-quality pseudo label data were obtained and served as input for the final mean teacher model training. The proposed methodology was validated on a dataset of 9069 synthetic images that correspond to five different subtypes of white blood cells. The model yielded an overall average accuracy of 94.4% with only 500 labeled samples, which is slightly lower than that of the fully supervised model with 9069 labeled samples (97.9%) but much higher than that of the fully supervised model with 500 labeled samples (86.5%). With such results, the proposed model demonstrates promising prospects for developing clinically useful solutions that are able to detect white blood cells based on blood cell images.