Abstract
The application of deep learning is rapidly transforming the field of bioimage analysis. While deep learning has shown great promise in complex microscopy tasks such as single-cell segmentation, the development of generalizable foundation deep learning segmentation models is hampered by the scarcity of large and diverse annotated datasets of cell images for training purposes. Generative Adversarial Networks (GANs) can generate realistic images that can potentially be easily used to train deep learning models without the generation of large manually annotated microscopy images. Here, we propose a customized CycleGAN architecture to train an enhanced cell segmentation model with limited annotated cell images, effectively addressing the challenge of paucity of annotated data in microscopy imaging. Our customized CycleGAN model can generate realistic synthetic images of cells with morphological details and nuances very similar to that of real images. This method not only increases the variability seen during training but also enhances the authenticity of synthetic samples, thereby enhancing the overall predictive accuracy and robustness of the cell segmentation model. Our experimental results show that our CycleGAN-based method significantly improves the performance of the segmentation model compared to conventional training techniques. Interestingly, we demonstrate that our model can extrapolate its knowledge by synthesizing imaging scenarios that were not seen during the training process. Our proposed customized CycleGAN method will accelerate the development of foundation models for cell segmentation in microscopy images.