Identification of white blood cells for the diagnosis of acute myeloid leukemia

Abstract

The diagnosis of hematologic malignancies such as leukemia includes identifying malignant white blood cells. Manual microscopic analysis of white blood cells is time-consuming and involves the assistance of medical specialists, and its accuracy may be affected by their abilities. Our objective is to develop an automated system that can identify 15 categories of white blood cells to assist medical practitioners in diagnosing acute myeloid leukemia. The proposed approach uses an ensemble model built with transfer learning-based pretrained networks to categorize 15 kinds of white blood cells. An over-sampling strategy is used to alleviate the problem of class imbalance, followed by data augmentation. For experimentation, a microscopic blood image dataset containing 18 365 cells images from 200 individuals representing 15 different forms of leukocytes is used. For cells with more than 400 images available, the suggested technique achieves an F1 score of more than 91%. Myeloblasts, which are frequent in myeloid leukemias and are detected in the peripheral blood, are recognized with an average precision of 95.74%. This work describes an image processing strategy that employs ensemble learning to assist in diagnosing acute myeloid leukemia by classifying 15 different types of leukocytes.Experiments show that our technique is both practical and effective. Extensive research supports the use of the leukocyte classifier in real-world medical applications, such as assisting clinicians in diagnosing diseases and decreasing human resource requirements.