Classificação automática de imagens de células sanguíneas usando redes neurais convolucionais

Abstract

This work aims to establish an automatic method for classifying blood images to support the diagnosis of diseases. The classification of white blood cells (WBC), which refers to the cells in the images used in this work, is an indicator of patient health that serves as the basis for several diagnoses. Currently, automated counters such as cytometers are used, which are expensive. Another alternative is manual counting on microscope slides. The database used consists of 12,500 enlarged images of blood cells, with the respective labels identifying each cell. In total, four types of blood cells are classified, with approximately 3,000 images for each class, therefore, a database considered balanced. The cell types are called Eosinophils, Lymphocytes, Monocytes and Neutrophils. Among the main tools available, the use of convolutional networks for intelligent image classification stands out. For this work, the CNN network architecture known as ResNet-18 was used, which demonstrated high accuracy in the classification of these images, with values greater than 95%. When implementing this network architecture, a parallel was established between its characteristics and its applicability in classification and technical feasibility of implementation for the task of classifying blood cell images were discussed. The results obtained exceeded the accuracies compared with previous works, obtained from the literature.