Robust Non-Invasive Support System for Blood Cell Image Classification Using Generative Adversarial Networks
Mohammad Shabaz, Mukesh Soni, Mohammed Wasim Bhatt · 2025
This chapter proposes a multi-scale conditional generative adversarial network using a multi-scale discriminator that generates numerous realistic leukocyte images. This approach can be used for classification and mitigate the issue of low detection accuracy attributed to limited leukocyte samples and imprecise features of the cell pictures. Blood cell pictures may be created, classified, and detected using the network training set approach. True and false discriminators, multi-scale convolution kernels, pooling domains, and channel splicing are introduced into the current generative confrontation network to boost the discriminator’s learning capabilities. Furthermore, the gradient similarity loss approach was developed to increase the final cell picture’s brightness and edge sharpness while preserving the image’s authenticity. Experiments revealed that incorporating a multi-scale discriminator and gradient similarity enhances the generated cell image quality, and that increasing the number of cell samples increases the cell classification and detection accuracy from 90.3% to 94.7% when compared to the case of actual data training. The classification and detection accuracy of cells is significantly improved when compared with the other models.