An Improved Unsupervised White Blood Cell Classification via Contrastive Learning
Abstract: The classification and counting of white blood cells (WBCs, leukocytes) in blood smears are of great significance for clinicopathological diagnosis. Therefore, the classification of WBCs in the images is a basic task. Most of the existing WBCs classification methods are based on supervised learning, which highly depends on a large number of image labels. To cope with the challenge of image annotation, in this paper, we propose an unsupervised WBCs classification method, which combines the advantages of contrastive learning and a deep clustering algorithm. Specifically, the proposed method firstly employs contrastive learning to pre-train the feature encoder, which is able to improve the similarity of feature coding among the same kind of cell categories. Then, the classical clustering algorithm is used based on the pre-trained image features for unsupervised classification of WBCs. Finally, the high confidence clustering results are fed back to the pre-trained feature encoder, which can be used as the pseudo labels to form a closed loop. Experimental results on a dataset containing 574 leukocyte images demonstrate the effectiveness of the proposed method.
Loading