Lightweight White Blood Cells Detection Using Fusion of YOLOv5 and Attention Model

Abstract

The human body is protected by an immune system which mainly consists of white blood cells (WBCs). There are five types of white blood cells, and each type will fight certain viruses and bacteria that are encountered in the human body. This defence system helps to maintain human health. Consequently, healthy WBCs keep humans healthy. Abnormality in WBCs can cause harmful viruses or bacterial infections. Leukaemia is a common WBCs disease which affects the production of good cells. Early detection is important for advanced treatment for cancer patient. One of the detection methods is by visual detection of the blood microscopic image since the five types of the WBCs are visually distinctive. In current practice, the pathologist will perform the diagnosis manually which may take time if there are many samples to examine. This procedure can be improved by automating it using a computer aided detection system. This paper studied the deep learning detection model of YOLOv5s and the effect of fusing the Squeeze-Excitation (SE) and Convolutional Block Attention Model (CBAM) into the YOLOv5s. It was performed on the four types of the WBCs, eosinophil, lymphocyte, monocyte, and the neutrophil taken from a public dataset. Based on the findings, the proposed method of YOLOv5s-SE, YOLOv5s-CBAM, and YOLOv5s-SE-CBAM produced overall accuracy of 99.5%, 99.5% and 99.4% mAP value and the performance are at par with the deeper model YOLOv5m with 65.8% of a smaller number of hyperparameters.