Synthesize More Diverse and Realistic Nanoparticle Image-mask Pairs on Limited Data

Automated biomedical image analysis, such as nanoparticle segmentation and phenotyping for drug discovery, requires robust learning models that are trained on large datasets of microscope images with appropriate annotations. However, current image synthesis techniques have limited performance in terms of diversity and fidelity, which can impact the accuracy of these models. Additionally, most learning-based image synthesis methods require a significant amount of training data for successful model convergence, resulting in training challenges on limited data. To address these limitations, we propose a layout-controllable conditional GAN with multi-level discriminators to synthesize more diverse and realistic homogeneous nanoparticle image-mask pairs with minimal guidance from training data. Our approach utilizes a variant image layout as the condition for the generator to vary the spatial distribution of nanoparticles in an image, generating image-mask pairs with more diversity. We also design image-level and object-level discriminators for multilevel discrimination, with loss constraints on images, masks and their correlations, ensuring high-fidelity and accurate synthesis. Experimental results demonstrate that the proposed network, trained on a small dataset, is capable of generating homogeneous nanoparticle image-mask pairs with sufficient diversity and fidelity.