Generalizability in White Blood Cells’ Classification Problem

Abstract

AbstractCounting and classifying white blood cells (WBCs) in blood samples helps the early diagnosis of the disease. Many works have been done to develop machine learning-based methods to count WBCs. However, most of these works have low generalizability, and their accuracy decreases sharply as the dataset changes. In this paper, a new method is presented that helps to increase the generalization power. In this method, first, the WBC’s nucleus is segmented, and then its convex hull is obtained. By subtracting the nucleus from the convex hull, a new image is created called the representative of the convex hull (ROC). Then, by Training a convolutional neural network (CNN) with the cells’ RGB image as well as the binary images of the nucleus and ROC, the generalization power is increased. The proposed method was first trained on the Raabin-WBC dataset, then its performance was evaluated on the LISC dataset without retraining. The proposed method’s accuracy on the Raabin-WBC and LISC datasets is 93.97% and 51.57 %, respectively. Besides, the generalization power of four well-known CNNs named VGG16, ResNext50, MobileNet-V2, and MnasNet1 was investigated. It was found that VGG16 has more generalization power among these models.