Establishment and application of peripheral blood leukocyte classification model

Objective To automatically classify peripheral blood leukocytes based on Swin Transformer model, and to compare the difference between Swin Transformer model and ResNet that is a classical convolutional neural network model.

Methods The classical convolutional neural network model ResNet and the new Swin Transformer model were used as network prototypes for training. White blood cell images were collected using the Cella Vision DI60 automatic analyzer, and the category labels of cells were confirmed by two experienced inspectors. The exponential attenuation, a learning rate attenuation method, was used to make the model converge faster. Then, 2, 788 leukocyte images were tested.

Results The average test accuracy of ResNet for five kinds of leukocyte images was 95.2%, while that of Swin Transformer was as high as 99.1%. Among them, the recognition accuracy of Swin Transformer model is 99.8% for neutrophils, 94.8% for eosinophils, 97.5% for basophils, 99.5% for lymphocytes and 93.8% for monocytes.

Conclusion Swin Transformer model reduces the amount of calculation, and is more suitable for leukocyte classification and recognition. Compared with ResNet, this model has more advantage in accuracy.