Abstract
In this work, we propose a novel GAN-based architecture, termed Physics-Inspired GAN (PI-GAN), to generate synthetic bimodal data comprising both intensity and phase images as produced through Fourier Ptychographic Microscopy (FPM). By explicitly incorporating the forward model of image formation into the GAN architecture, our approach ensures that the physical relationship between the intensity and phase modalities is preserved throughout the training and generation processes, therefore solving the mode collapse problem encountered in classical GANs. Our approach is evaluated for the classification of the five major types of white blood cells (WBCs) in peripheral blood smears, a domain where severe class imbalance is a major challenge. In particular, basophils represent less than 1% of circulating WBCs, making it difficult to train robust classifiers without synthetic augmentation. To overcome the scarcity of basophil data, we proposed a two-step fine-tuning strategy: first training the PI-GAN to generate neutrophils (a more abundant but morphologically similar class), and then adapting the model to produce basophils. Our results show that the addition of synthetic basophil images allows a great improvement (5% in precision) in the ability to correctly classify basophils. Our approach offers great potential for future hybrid models that combine physics-based priors with the flexibility of deep generative networks.