MinimalGAN: diverse medical image synthesis for data augmentation using minimal training data

Image synthesis techniques have limited application in the medical field due to unsatisfactory authenticity and precision. Additionally, synthesizing diverse outputs is challenging when the training data are insufficient, as in many medical datasets. In this work, we propose an image-to-image network named the Minimal Generative Adversarial Network (MinimalGAN), to synthesize annotated, accurate, and diverse medical images with minimal training data. The primary concept is to make full use of the internal information of the image and decouple the style from the content by separating them in the self-coding process. After that, the generator is compelled to concentrate on content detail and style separately to synthesize diverse and high-precision images. The proposed MinimalGAN includes two image synthesis techniques; the first is style transfer. We synthesized a stylized retinal fundus dataset. The style transfer deception rate is much higher than that of traditional style transfer methods. The blood vessel segmentation performance increased when only using synthetic data. The other image synthesis technique is target variation. Unlike the traditional translation, rotation, and scaling on the whole image, this approach only performs the above operations on the segmented target being annotated. Experiments demonstrate that segmentation performance improved after utilizing synthetic data.