Retinal Recombination: A robust augmentation and synthesis of retinal images with annotation fusion

Medical image synthesis has become a crucial research topic for enhancing image analysis software. The relatively small size of available datasets, the lack of ground-truth annotations, and the poor quality of registered images are among the driving reasons for the development of novel algorithms. Recently, neural network architectures for image generation, such as generative adversarial networks, have achieved high fidelity in this domain. However, the remaining problems include high computational cost, low output resolution, and inconsistent synthetic features. To address these limitations, we propose a novel retinal image augmentation and synthesis algorithm that decomposes real images, augments their components, and recombines them into new, unique samples. Further, we introduce a method for optic disc transfer based on image statistics and blood vessel geometry augmentation based on entropy maximisation. Our approach works at a scalable resolution on standard personal computers, transfers the annotations from the original images, and enables the fusion of features from labelled and unlabelled datasets. Two ophthalmologists evaluated the synthetic image quality, and we validated our method using four publicly available datasets, showing the ability to preserve segmentation masks, statistically significant improvements in optic nerve head segmentation, and the impact on pathology classification. Our results show potential for high-resolution image synthesis and for improving the automatic detection of retinal features essential for computer-aided clinical diagnosis.