RDNet: ResNet-18 with Dropout for Blood Cell Classification

(Aims) Blood cells are hematopoietic pluripotent stem cells derived from bone marrow. Blood diseases occur primarily in the hematopoietic system and can affect the hematopoietic system with abnormal blood changes, characterized by anemia, bleeding, and fever. It is helpful for doctors to diagnose blood diseases by classifying blood cells. However, doctors take a lot of time and energy to classify blood cells. The classification process is easily disturbed by external factors, such as doctors’ lack of rest, fatigue, etc. Many researchers use CNN to classify and detect red blood cells or white blood cells. However, using CNN has some problems in the classification or detection process. First, most researchers only classify blood cells into two categories, but there are many different types of blood cells. In addition, some studies are multi-classification of cells, but the results are often not ideal. (Methods) We propose a new model (RDNet) for the automatic classification of four types of blood cells to deal with these problems. The proposed RDNet selects the pre-trained ResNet-18 as the backbone. We transfer the pre-trained ResNet-18 because of the difference between the blood cell data set with the ImageNet data set. We add dropout to improve the classification performance. (Results) The accuracy of the proposed RDNet is 86.53%. The proposed RDNet obtains better accuracy than the transferred ResNet-18 because we add dropout in RDNet. Based on the accuracy, the proposed model is an effective tool to classify blood cells