Conditional generation of medical images via disentangled adversarial inference

Synthetic medical image generation has a huge potential for improving healthcare through many applications, from data augmentation training machine learning systems to preserving patient privacy. Conditional Adversarial Generative Networks (cGANs) use conditioning factor generate images and have shown great success in recent years. Intuitively, the information an can be divided into two parts: 1) content which is presented vector 2) style undiscovered missing vector. Current practices using cGANs generation, only single variable (i.e., content) therefore, do not provide much flexibility nor control over generated image. In this work we propose DRAI—a dual adversarial inference framework with augmented disentanglement constraints—to learn itself, disentangled representations of content, impose process. framework, learned fully unsupervised manner, while both supervised (using vector) (with mechanism). We undergo novel regularization steps ensure content-style disentanglement. First, minimize shared between by introducing application gradient reverse layer (GRL); second, introduce self-supervised method further separate variables. For evaluation, consider types baselines: latent models that infer variable, double variables (style content). conduct extensive qualitative quantitative assessments on publicly available imaging datasets (LIDC HAM10000) test conditional retrieval style-content show general, achieve better performance give more also our proposed model (DRAI) achieves best score overall performance.