Identification and counting white blood cell subtypes with convolutional neural network

Abstract

White blood cell (WBC) has five subtypes namely neutrophil, eosinophil, basophil, lymphocyte, and monocyte which play specific roles in the immune system and against diseases. The object detection model applied to microscopic objects is introduced to assist experts in performing tasks in blood analysis. Unbalanced cell composition of WBC subtypes to be detected is a challenge in building a model in Convolutional Neural Network (CNN). This research aims to build models in recognizing and counting WBC subtypes with neural networks constructed from augmented data enrichment. CNN is demonstrated in this study with the YOLOv5s, YOLOv5l, and YOLOv5x models to detect and count WBC subtypes. Generating three different datasets, the first is raw data with a limited and unbalanced amount, the second dataset is augmented data with geometric operation based, and the third dataset is augmented data with image enhancement based. The experimental results show that in recognition and counting systems for the five subtypes of WBC, the best model among the three models of the YOLOv5 family is YOLOv5l. The best accuracy among built models is YOLOv5l from the augmentation with image enhancement based which has an accuracy of Mean Average Precision (mAP) mAP@.5 0.995 and 0.988 mAP@.5:.95 at 600 epochs of training.