Implementation and Improvement of EfficientNetV2 Model With GRU to Classify White Blood Cell Types

White blood cells are formed in the bone marrow which is responsible for the immune system, protecting the body from disease, and infections and other foreign objects. In various diseases such as leukemia, anemia, malaria, and human immunodeficiency virus (HIV) infection can be predicted by observing the number of white blood cells formed. This information highlights the importance of understanding the type of white blood cell content as a tool in the disease diagnosis process. Various studies have adopted a deep learning approach to analyze white blood cell images, one of which is EfficientNetV2. Although EfficientNetV2 has been designed to increase efficiency in the training process, the use of large image sizes can still slow down the training process so that the initial layers in the model become less stable. The addition of GRU can improve the model's ability to understand and model relationships in the features produced by the model. In this study, the EfficientNetV2 model is combined with GRU to improve the model's performance in analyzing and classifying white blood cell images. The Blood Cell Counting and Detection (BCCD) dataset consisting of four classes, namely eosinophil, neutrophil, monocyte, and lymphocyte will be used during testing. This study involves two experimental scenarios on the dataset to see the difference in performance between the two approaches. The results of the average metric values in the first scenario obtained a loss of 0.3590, accuracy 0.9455, precision 0.9531, recall 0.9531, and f1 score 0.9531. Then in the second scenario obtained the results of the average loss value of 0.0091, accuracy 0.9977, precision 0.9978, recall 0.9977, and f1 score 0.9977. These results show the ability of EfficientNetV2 with GRU which is relatively good in classifying images efficiently and accurately on the BCCD dataset.