Most modern segmentation techniques for ocular images are based on deep learning methods and are thus critically dependent on large-scale annotated datasets. Unfortunately, suitable datasets are labour-intensive to gather and often raise privacy concerns. To address these issues, we present a novel framework, called BiOcularGAN, capable of generating large-scale synthetic datasets of photorealistic ocular images, in both the visible and the near-infrared light spectrum, along with corresponding segmentation masks. The framework is centered around an innovative Dual-Branch StyleGAN2 model, which facilitates the generation of high-quality aligned bimodal images. By exploiting latent features of the model, the framework is also able to produce extremely accurate segmentation masks of the synthetic images, based only on a handful of manually labeled examples, therefore minimizing human involvement. We evaluate the BiOcularGAN framework through extensive experiments across five diverse ocular datasets and analyze how bimodal data generation affects the quality of produced images and masks. In addition, we showcase that the generated data can be employed to train highly successful deep segmentation models, which can generalize well to other real-world datasets.
|
