Fundus Image Enhancement via Semi-Supervised GAN and Anatomical Structure Preservation

The fundus image is often used in clinical diagnosis due to ease and safety of acquisition, but the quality may be affected by environment and onsite operations. As low-quality medical images may lead to misinterpretation in diagnosis and analysis, it is important to improve quality of the improperly-acquired fundus images. Unfortunately, the existing fundus image enhancement methods require task-specific prior knowledge or suffer from insufficient generalization ability. To cope with this issue, a generative adversarial network (GAN) based model is proposed, namely the semi-supervised GAN with anatomical structure preservation (SSGAN-ASP). Specifically, an anatomical structure extraction component is employed in the generator to guide the enhancement process by preserving both retinal and lesion structures, while color information in the fundus image is also preserved. The SSGAN-ASP model is evaluated and compared with the state-of-the-art methods for medical image enhancement on three popular datasets. In addition, it is applied in the pre-processing of retinal vessel segmentation and diabetic retinopathy grading tasks to show efficacy in computer-aided diagnosis. Experimental results demonstrate that visual quality of the enhanced image can be improved while better performance in clinical diagnosis is achieved with our proposed model by adopting the anatomical structure extraction component and preserving color information as well.

I. Introduction

The fundus image plays a crucial role in screening and diagnosing ocular diseases, such as diabetic retinopathy, age-related macular degeneration, and glaucoma [1]. Due to the limitations of acquisition environment, imaging equipment and operator's experience, the acquired fundus images often tend to exhibit unstable quality. A screening study by U.K. BioBank reported that about 30% of fundus images are lack of adequate quality for clinical diagnosis [2] or automatic fundus image analysis applications (e.g., retinal vessels segmentation [3], optic disc/cup detection [4] and diabetic retinopathy grading [5]). The low-quality fundus images often contain dark, over-exposure, uneven light, blur, artifact or color distortion. As high-quality fundus images have clear retinal structures, it is hard to observe the complete structure of anatomical tissues in low-quality images, including optic disk/cup, retinal blood vessels, fovea and lesion regions. Thus, improving the quality of improperly-acquired fundus images for better visibility is still a challenging problem.