Fine-grained imbalanced leukocyte classification with global-local attention transformer

Abstract

Leukemia is a fatal disease that requires the counting of White Blood Cells (WBCs) in bone marrow for diagnosis. However, bone marrow blood contains many types of leukocytes, some of which have subtle differences. To address this issue, we propose the WBC-GLAformer model, which comprises three parts: Low-level Feature Extractor (LFE), Global–Local Attention based Encoder (GLAE), and Discrimination Part Select (DPS). The LFE uses a convolutional neural network (CNN) to tokenize patches from the extracted low-level features. The GLAE combines the ability of the CNN to extract local features with the ability of the transformer to extract global features, thereby enriching the features of leukocyte images. The DPS improves the accuracy of leukocyte classification by selecting the discriminative regions. Our method achieves state-of-the-art results in the bone marrow leukocyte fine-grained classification dataset. Experimental results demonstrate that the model has good generalization on different datasets and is more robust to the optimizer. And visualization results show that the model can effectively focus on the discriminative parts of different cells. Code is available at https://github.com/ywj1/WBC-GLAformer