Abstract:
Data augmentation addresses the critical challenge of limited data in medical imaging. While generative adversarial networks (GANs) have been a popular choice in synthesizing medical images, controlled generation targeting disease-specific semantic has been difficult, partly due to the difficulty to disentangle local disease-specific semantic factors from global disease-irrelevant factors. In this work, we present a semantic image editing framework for medical image augmentation that is able to generate smooth variations along the desired direction of disease attributes in user-defined regions of interest. This is achieved by discovering the optimal trajectory on the latent manifold of a pre-trained StyleGAN, guided by a mask of the region of interest and explicitly constrained by desired directions of semantic changes. We test the presented method on the public Chest X-ray dataset. To evaluate the quality of the generated medical images, we leverage both domain experts (pulmonologists) for qualitative assessments and present a novel metric to quantify the ability of the presented method to generate progression of disease severity in the synthesized images. We also show that data augmentation using the presented method improves downstream classification tasks.
