Siamese Network-Based Few-Shot Learning for Classification of Human Peripheral Blood Leukocyte

In the present study, we propose a Siamese network-based few-shot learning method to classify rare human peripheral blood leukocyte images. The proposed Siamese network contains two identical convolutional neural networks (CNNs) and a logistic regression network (LRN). The CNNs are utilized to extract image features and the LRN is responsible for comparing the similarity of these features. To train the Siamese network, two leukocyte samples that belong to the same category make up genuine pairs and two leukocyte samples from disparate categories make up impostor pairs. A strategy for generating sample pairs is designed for this study to ensure the number of genuine pairs and impostor pairs is equal. After the Siamese network accomplishes the training process, with the purpose of evaluating the classification performance of the proposed Siamese network, an appropriate support set is selected to test the samples in the query. The numerical results show that the Siamese network can overcome the scarcity and imbalance of datasets used in this research, and the average classification accuracy of rare leukocyte images can achieve 89.7%, which is potential to apply the Siamese network-based few-shot learning method for addressing the issue of rare leukocyte images recognition in medicine.