Image enhancement of wide-field retinal optical coherence tomography angiography by super-resolution angiogram reconstruction generative adversarial network

Wide-field retinal optical coherence tomography angiography (OCTA) usually suffers from low image resolution in clinical practice because of insufficient lateral sampling. In this study, we develop a deep-learning-based method named super-resolution angiogram reconstruction generative adversarial network (SAR-GAN) to enhance the en face OCTA image quality. A sophisticated home-made spectral-domain OCTA system is employed to capture the data of retinal angiograms with different scanning protocols. High-resolution 3 × 3 mm2 OCTA images and low-resolution (LR) 6 × 6 mm2 OCTA images are utilised in training the network. We propose an improved loss function for SAR-GAN for the reconstruction of perceptually enhanced super-resolution images. The well-trained network is utilized to processing the LR OCTA images with a field of view (FOV) of 3 × 3 mm2, 6 × 6 mm2 and as large as 9 × 9 mm2. The qualitative and quantitative comparisons show that SAR-GAN provides perceptually better visualization and significantly enhances the image quality in terms of noise intensity, contrast-to-noise ratio and vessel connectivity. Moreover, it demonstrates superior image enhancement for retinal OCTA with small or large FOVs, compared with other traditional and deep-learning based methods. The SAR-GAN has great potential to improve the clinical assessment by wide-field OCTA.