Leukocyte detection in blood smear image based on improved YOLOv3

Abstract

The classification of leukocytes is an important indicator for the detection of a variety of blood diseases in routine blood tests. Although there have been many papers studying the detection of WBCs (white blood cell) or classification independently, few papers consider them together. In this paper, we propose an end-to-end white blood cell localization and classification method based on improved YOLOv3. Firstly, we utilize the k-means clustering algorithm to generate the anchor boxes suitable for WBC. Secondly, to identify white blood cells of different sizes, we use multi-scale predictions with YOLOv3 network structure. Experimental results on both the LISC dataset [1] and a dataset of 7500[2] leukocyte smear images of five categories demonstrate that the proposed improved-YOLOv3 can achieve efficient detection performance in terms of accuracy. The recognition accuracy of the two data sets is reached respectively 96.4% and 95.5%.