Machine Learning Approaches for Improving the Accuracy of Blood Cell Detection and Subtypes Classification Using Smear Microscopic Images

The human blood contains red blood cells (RBC), white blood cells (WBC), platelets, and plasma. The entire blood cell count defines the state of health. A normal human has RBCs ranging from 4.5 to 6.0 million cells per microliter in males and 4.0–5.0 million cells per microliter in females, as well as WBCs ranging from 4.5 to 11.0 thousand cells per microliter in both males and females. The segmentation and identification of blood cells are extremely vital. The RBC and WBC counts are extremely important to diagnose varied diseases such as haemolytic anaemia, nutritional anaemias, acute myeloid leukaemia, chronic myelogenous leukaemia, and chronic lymphocytic leukaemia. The count of blood cells is performed in manual hospital laboratories using a victimisation device known as a hemocytometer and magnifier. However, this technique is tedious, laborious, and time-consuming, and it produces incorrect results due to human error. Also, there are some expensive machines, like instruments, that do not seem to be reasonable in each laboratory. The proposed method has involved the use of image processing to classify the blood cells with the help of ResNet deep neural networks. The algorithm can extract the feature of each segmented cell image and classify the types. The overall accuracy was 93.01%. The system has been developed to provide accurate and fast results using large dataset of blood smear images.