Classification of peripheral blood cell images using deep learning

Abstract

Diagnosis and Identification of cells and disease infected cells are and important part of in medical science that bears huge significance even today. There are health implications can often be identified my observing the morphological changes of cells as well as the quantity of cells. The traditional methods of counting blood and chemically identifying diseases can be expensive and/or time consuming to the extent that only certain medical centres can perform the task at hand, or take days to receive a report of. However, we believe Deep Learning with Convolutional Neural Networks (CNNs) can take over most of this tedious process. In this work, we aim towards creating a custom CNN model that can quickly classify different kinds of peripheral blood cells such as the 7 different white blood cell types (basophils, erythroblasts, ig, eosinophils, lymphocytes, monocytes, neutrophils) and platelets. Such a model can be used in blood cell counts which can be used to identify cases like leukemia. Moreover, such a method can be extended into other fields such as red blood cell detection or even infected cell detection, which includes identifying diseases from Sickle Cell Anemia to cells affected by Covid19. Our custom CNN model has performed exceptionally well, achieving accuracies as high as 99.1% and 98.9% in training and validation respectively, which is significantly higher than using pre-trained models such as DenseNet or NasNet. In more ways than one, we show how our model is better suited for the task at hand.