White blood cells classification of peripheral blood smear based on machine learning

Abstract

In clinical practice, peripheral blood smears are observed with the naked eye to confirm the number and morphology of White Blood Cells (WBC). The pathologist’s proficiency has a great influence on the results because WBCs have diverse morphology. Many studies have been conducted on diagnostic assistance system for pathologists. However, these are still difficult to apply to images of various extracellular environments that change according to acquired conditions. Added to this is the limitation that only one WBC can be classified in one image. Therefore, in this study, we propose a robust segmentation algorithm for WBC nucleus and cytoplasm used color space, superpixel and watershed algorithm. Further, we propose a classification algorithm that classifies five types of normal WBCs using 16 morphological and texture features as well as the K-Nearest Neighbor (KNN) model. The accuracy of segmentation WBC nucleus and cytoplasm was 95.83% and 93.66%, respectively. The p-value for all the 16 features was significant (<0.001). In addition, through this study, research on a diagnostic assistance system that can classify more types of WBCs will be possible in the future.