DSIS: Investigating retinal image synthesis with dual-stream super-resolution learning

Deep learning-based medical image processing is an important technique for computer-aided diagnosis of fundus images. However, patient privacy concerns and the high cost of labeled data often lead to a scarcity of labeled data, which limits the development of deep learning methods in this field. To address these issues, this paper explores dual-stream image synthesis (DSIS), a method that converts semantic segmentation maps into retinal fundus images. The proposed method consists of the image synthesis module, the single image super-resolution module, and the multiscale feature extraction module. Specifically, we aim to utilize single-image super-resolution as an auxiliary task to enhance the quality of synthesized retinal fundus images without introducing additional computational costs. Additionally, the multiscale residual attention strategy is applied to capture information at different levels and improve the perception of fine structures in the images. The experiments were conducted to demonstrate the effectiveness of our method in generating retinal images. The obtained results demonstrate the ability of our approach to produce visually captivating and anatomically precise retinal fundus images, confirming its potential for a wide range of applications in retinal image analysis.