Publication

Abstract : Diabetic Retinopathy (DR) is considered one of the most serious causes of visual impairment and permanent blindness in the world, causing damage to the optic nerve head without any prior clinical sign. Colored retinal images are the most often utilized imaging data in the medical sector for screening and diagnosing eye disorders, and fundus cameras are commonly employed in clinical research to collect these images. Due to a number of reasons, such as exposure discomfort, incorrect equipment parameter setting, improper operation, and a lack of proficiency on the part of various medical professionals during the image-acquisition process, there are many poor-quality retinal scans in the current database. The main purpose of this research is to perform early detection of DR by examining different levels of Proliferative Diabetic Retinopathy (PDR) and Nonproliferative Retinopathy (NPDR) images using deep learning algorithms. This work recommends employing EiGAN (EnlightenGenerative Adversarial Network) to enhance the quality of the fundus images. The ROI-based segmentation and subsequent classification of the quality-enhanced fundus images are accomplished with the help of the proposed Convolutional Transfer Learning Neural Network (CTLNN). The dataset was obtained from Precious Optics, Mysuru, India. The image dataset contains 1920 samples of PDR and NPDR fundus retinal images. Among these, there are multiple classes under NPDR, such as level-1 (healthy), level-2 (mild), level-3 (moderate), and level-4(severe), respectively. This research shows that the methods such as transfer learning and Efficient Net-B7 design to achieve robustness and efficiency. The methodology observes a significant increase across a range of performance indicators through a review of pre and post-enhancement findings.