An Unsupervised Fundus Image Enhancement Method with Multi-Scale Transformer and Unreferenced Loss

Abstract

Color fundus images are now widely used in computer-aided analysis systems for ophthalmic diseases. However, fundus imaging can be affected by human, environmental, and equipment factors, which may result in low-quality images. Such quality fundus images will interfere with computer-aided diagnosis. Existing methods for enhancing low-quality fundus images focus more on the overall visualization of the image rather than capturing pathological and structural features at the finer scales of the fundus image sufficiently. In this paper, we design an unsupervised method that integrates a multi-scale feature fusion transformer and an unreferenced loss function. Due to the loss of microscale features caused by unpaired training, we construct the Global Feature Extraction Module (GFEM), a combination of convolution blocks and residual Swin Transformer modules, to achieve the extraction of feature information at different levels while reducing computational costs. To improve the blurring of image details caused by deep unsupervised networks, we define unreferenced loss functions that improve the model’s ability to suppress edge sharpness degradation. In addition, uneven light distribution can also affect image quality, so we use an a priori luminance-based attention mechanism to improve low-quality image illumination unevenness. On the public dataset, we achieve an improvement of 0.88 dB in PSNR and 0.024 in SSIM compared to the state-of-the-art methods. Experiment results show that our method outperforms other deep learning methods in terms of vascular continuity and preservation of fine pathological features. Such a framework may have potential medical applications.