A hybrid approach for efficient multi-classification of white blood cells based on transfer learning techniques and traditional machine learning methods

Abstract

Diagnosed blood-related diseases include the identification of blood samples taken from the patient. Therefore, the classification of white blood cells, also known as leukocytes, is substantial for differentiating leukemia and blood diseases. In this article, we aimed to classify four different types of white blood cells. Two different transfer learning methods are applied to achieve this goal. In the first method, AlexNet, ResNet18, and GoogleNet architectures are retrained with the fine-tuning method using the dataset we have provided from the Kaggle and then given to classify on the softmax and SVM methods. With this method, the hybrid architecture where the ResNet18 is used with SVM achieves 99.83%.In the second method, feature transfer, the same architectures are implemented for feature extractors. First, extracted features of architectures are given to a variety of classifiers. Second, the features are taken from the architectures, and concatenated as pairs and triples, used to obtain 4 different feature sets consisting of 2000 and 3000 features. After that, these features are also subjected to the same classifiers. Finally, results are revealed by using the same classifiers after the PCA algorithm. The results illustrate that the proposed methods significantly contribute to white blood cell multi-classification.