In view of the uneven distribution of image categories and complex background information in acute lymphoblastic leukemia, as well as the challenges of time-consuming manual diagnosis and susceptibility to subjective factors, an EfficientNet-DSP leukemia classification method has thus been proposed. The generalization ability of the model can be enhanced by the proposed method through image enhancement techniques and dynamic random deactivation blocks, with residual permutation attention mechanism integrated to enhance its ability to extract detailed features. It is proposed to use Dy-ODConv dynamic convolution to learn information from various dimensions, and dynamically adjust the weights of convolution kernels, thus improving classification accuracy while reducing the number of parameters. In addition, the loss function of the algorithm has been improved to enhance the classification ability of the model when processing complex background images. Finally, experiments are conducted on the Blood Cells Cancer dataset, with the results showing that EfficientNet-DSP achieves an image classification accuracy of 98.46%, an improvement of 2.54% compared to the original EfficientNetV2 model, and an improvement of 3.61% compared to the optimal values of other algorithms. It can be concluded that the proposed method effectively improves the diagnostic accuracy of acute lymphoblastic leukemia images, which makes it a reference for physician diagnosis.