Deep Learning based Real-Time Noise Removal in Retinal Fundus Images

Real-time noise removal in fundus images, which are images of the back of the eye (retina), is a critical task in various medical applications, particularly in ophthalmology. This study presents a deep learning approach for denoising retinal fundus images affected by gaussian and speckle noise. A diverse dataset comprising various retinal image acquisition conditions was utilized, with synthetic speckle and gaussian noise introduced at different levels to simulate in real-world scenarios. In the proposed approach, the noise is introduced by considering the variance of 0.5 and 1. Moreover, pre-processing steps were employed to mitigate noise impact, followed by feeding the images into a proposed CNN-based denoising architecture. The architecture’s performance was rigorously evaluated using quantitative metrics and expert clinician assessments. Results include an average Peak Signal-to-Noise Ratio (PSNR) of 47.5 dB, an average Structural Similarity Index Measure (SSIM) of 0.90, and an average $R 2$ score of $71 \pm 0.7 \%$. Additionally, the BRISQUE score averaged 59.41. These results demonstrate the efficacy of the approach in enhancing image quality and preserving essential features, highlighting its potential for clinical applications in retinal imaging.