Efficient Diabetic Retinopathy Segmentation and Classification using Residual Encoder-Decoder Architecture

Diabetic retinopathy (DR), a severe complication arising from diabetes, poses a significant threat to vision due to the deterioration of retinal vessels. Recent techniques in DR detection, such as Convolutional Neural Networks (CNNs) and deep learning models, have shown promise but face challenges in accurately segmenting and classifying retinal images due to variations in image quality, occlusions, and the need for large annotated datasets. This study presents an innovative methodology for automated detection, grading, and segmentation of DR using deep learning, with a focus on residual encoder-decoder architecture. The study utilizes the Indian Diabetic Retinopathy Image Dataset (IDRID), comprising 81 fundus images and labels, to rigorously evaluate the proposed methodology. By employing advanced image preprocessing techniques to enhance data quality, followed by a unified model capable of both segmentation and classification tasks, the proposed method achieves competitive performance metrics. Specifically, the model demonstrates an accuracy of 85.2% and specificity of 86.1 % in segmenting and classifying DR features. These findings contribute to the improvement of diagnostic accuracy and patient outcomes in retinal diseases, offering potential applications in clinical settings to support early diagnosis and management of DR, thereby enhancing patient care and alleviating healthcare system burdens.