DCGAN-generated Synthetic Images Effect on White Blood Cell Classification

Abstract

White blood cell can give information about someone’s health. Imbalanced amount of white blood cells indicates someone’s disease. The disease detection using white blood cell can be done by using flow cytometry, however the use of the device has many drawbacks. The drawbacks can be solved by using computer aided classification. However, another problem arises, that the available white blood cell dataset has small dataset and the class distribution is imbalanced. This problem can be solved by using classic data augmentation and DCGAN to create synthetic image in order to balance the amount of white blood cell dataset. The balanced dataset then trained using ResNet50 model as the classification model. Accuracy, precision, recall and Fl-score are used as the performance measures of the classification. The result shows that the model obtain accuracy of 82.5%. ResNet50 was chosen because it can produce good accuracy when used in medical image classification.