A Modified Deep Neural Network-based Denoiser for Pre-Processing of Diabetic Retinopathy Images

Diabetic retinopathy (DR) is a significant reason for visual impairment all over the planet. Recognizing symptoms in the fundus picture is commonly utilized in illnesses connected with the eyes like diabetic retinopathy. It requires early detection through high-quality retinal imaging. However, noise and poor contrast degrade image clarity, affecting diagnosis. To overcome these limitations, pre-processing stage plays a vital role in clinical picture handling to build the nature of fundus images. Preprocessing technique is principally used to eliminate undesirable noises and improve some image features. Image processing makes use of a variety of pre-processing methods. This study presents an exhaustive preprocessing CAD model for diabetic retinal images. By harnessing both advanced deep learning methods and conventional image processing methods, this mechanized conclusion model (computer aided design) looks to work with better analysis and the executives of diabetic retinopathy - one of the main sources of vision misfortune all around the world. We propose a novel preprocessing pipeline integrating Wiener filtering, modified CLAHE, and a deep learning-based Retinal_Denoiser to enhance DR image quality. The proposed Retinal_Denoiser is utilized during the preprocessing phase of the computer-aided design model to eliminate noise and improve quality in retinal images, utilizing deep learning-based denoising autoencoder and protecting the fundamental features of DR pictures. Preprocessing steps were utilized to increase signal-to-noise ratio, our proposed model sets another best-in-class result with a peak signal-to noise ratio (PSNR) value of 62. The high PSNR achieved by our proposed method indicates that it is more effective in preserving the image details while effectively suppressing noise. The fundamental goal of this research is to improve the image by reducing noise, improving contrast, and preserving important structures such as blood vessels and the optical disc. Our method achieves a PSNR of 62.12, surpassing conventional CNN-based, RNN-based, DnCNN-based, and other well-known denoisers.