A Method for Classification and Detection of Blood Cell Images Based on a Multi-Scale Conditional Generative Adversarial Network

A conditional generative adversarial network (cGAN) structure utilizing a multi-scale discriminator was proposed to address the issues of low detection accuracy caused by insufficient leukocyte samples and unclear details in generated cell images. This approach aims to generate a large number of realistic leukocyte images, which are then incorporated into the training set of a classification detection network, thus facilitating the generation and classification of blood cell images. The proposed method introduces a multi-scale convolution kernel and a multi-scale pooling domain into the authenticity discriminator of the generative adversarial network, along with a channel connection to enhance the discriminator's ability to distinguish both micro-detail texture features and macro-geometric features. A gradient similarity loss function is also introduced to improve the brightness and edge clarity of the generated cell images, thereby enhancing their authenticity. Results demonstrate that the quality of the cell images is significantly improved with the inclusion of the multi-scale discriminator and the gradient similarity loss function during the image generation stage. Compared to training solely with real data, the addition of diverse cell samples in the image classification and detection stage leads to an increase in the average accuracy of cell classification and detection from 90.4% to 94.7%.