Retinopathy Detection using Retinal Fundus Images among Suspected Individuals

Abstract

Retinopathy is a condition represented by retinal degeneration due to diseases like diabetes and hypertension, which is also a leading cause of vision loss globally. The early detection and management of retinopathy are crucial for preventing severe vision impairment and blindness. Retinal fundus imaging, a non-invasive and widely used diagnostic tool, provides a detailed view of the retina, allowing for the identification of pathological changes associated with retinopathy. However, manual examination of these images by ophthalmologists is time-consuming, labor-intensive, and subject to variability, particularly in large-scale screening programs. newlineThis research introduces a cutting-edge approach for the automated detection of retinopathy in individuals suspected of having the condition, using retinal fundus images. The study leverages deep learning algorithms to develop a robust system capable of accurately identifying early signs of both common and rare ocular diseases, thereby enhancing the accuracy and reliability of retinal disease detection. The study introduces a unified framework for automated retinopathy analysis using retinal fundus images organized into three sequential phases. newlinePhase 1 applied Convolutional Neural Networks (CNN) to grade diabetic retinopathy, achieving 89% accuracy, exceeding previously reported CNN-based models that peaked at 82%. This stage established a reliable baseline for disease severity assessment. newlinePhase 2 expanded the scope to multiclass classification of retinal disorders via transfer learning. Adapting the VGG19 architecture on the Retinal Fundus Multi-disease Image Dataset (RFMiD) yielded 94% accuracy in a two-class scenario. A customized EfficientNetB0 provided an optimal balance between accuracy and computational load, while a tailored DenseNet121 delivered the highest precision for complex patterns. When compared against other state-of-the-art models, accuracies ranged from 87% to 91%. newlinePhase 3 focused on segmenting critical retinal structures, exudates and the optic disc using a specialized U Net model with binary cross entropy loss and the Adam optimizer. The model attained an Intersection over Union (IoU) score of 1.00 and a validation accuracy of 1.00. Integration of Otsu thresholding further refined boundary newlinedelineation. Earlier segmentation approaches employing transformer-based and encoder decoder networks reported Dice scores near 0.88 and IoU values up to 0.84. newlineCollectively, the outcomes of the three phases demonstrate the effectiveness of deep CNNs for grading, transfer learning for multiclass classification, and U Net segmentation in automated retinopathy analysis. The proposed system offers enhanced diagnostic precision and holds promise for reducing clinical workload by enabling prompt and accurate retinopathy screening and diagnosis. newlineKeywords: Automated Diagnosis, Deep Learning, Diabetic Retinopathy, Fundus Imaging, Retinal Image Classification, Transfer Learning, U-Net Segmentation newline newline newline newline newline