An Improved EfficientFormerv2 Network for Bone Marrow Cell Classification

Classification and counting of bone marrow cells is necessary for the diagnosis and treatment of various blood disorders. However, this job needs to be done manually by doctors, which is a long period and high workload. In order to realize the automatic identification of bone marrow blood cells, this paper proposes a model structure based on the improved Efficient Formerv2, which optimizes the token mixer in the model mixing block, and the activation function between convolutions in the feedforward neural network is upgraded from Gelu to Mish. Furthermore, addressing the potential issue of gradient explosion caused by the loss function, the loss function is modified from Cross Entropy to Focal Loss. The proposed model was tested on a public dataset of bone marrow cell images from the Munich Leukemia Laboratory, achieving a maximum classification accuracy of 94.85% in a six-class experiment. This research offers a promising technique for the automatic recognition and classification of bone marrow cell images.