Leukovit: An efficient vision transformer-based model for automatic classification of leukocytes

The identification and evaluation of leukocytes are important to assess the quality of the human immune system; however, the analysis of blood smears depends on the pathologist’s expertise. The manual method for analyzing and classifying WBCs is costly and time-consuming and can result in errors in detection. Most deep learning methods use CNN-based models for white blood cell classification. This paper discusses the use of a ViT-based network, for the classification of leukocytes (WBCs) in a blood sample. The Dataset used in this paper consists of 352 images with a size of 320x240, which was augmented through techniques to create a balanced dataset of 12444 images. The augmented data was then used to train a ViT-based architecture to classify the different types of WBCs. As the first step of the proposed algorithm, a convolutional tokenizer has been applied for patch extraction of images. These patches have been flattened and have been used as input for a ViT-based structure to recognize the subclasses in the second step. The results obtained using Leukovit show that the accuracy of the proposed network is 99.04% which is outperforming the state-of-the-art networks.