Enhanced Fundus Image Analysis for Retinal Disease Screening Using Deep Learning Framework

Abstract

ABSTRACT In this paper, a unified deep convolutional framework for multimodal ocular image-based diagnosis of three vision-threatening eye diseases, diabetic retinopathy (DR), glaucoma and age-related macular degeneration (AMD) is proposed. It takes retinal fundus photographs and slit lamp images (and an optional patient metadata) and defines a pipeline of resizing, contrast enhancement using Contrast-Limited Adaptive Histogram Equalization (CLAHE), normalization and augmentation before transformer based feature extraction. The proposed configuration pairs an EfficientNet-B3 fundus backbone with an EfficientNet-B0 slit lamp images backbone and a fully connected metadata network, fused by feature concatenation and directed by an adaptive disease router to specialist DR, glaucoma and AMD classification heads; ResNet, DenseNet and MedNet are retained as candidate baselines for comparison. This is a methodology and framework paper, and it outlines the datasets, preprocessing, architectures, training protocol and evaluation plan (including confusion matrices, and ROC and precision-recall curves) that would be used in a subsequent implementation to produce and report empirical results. No model has been trained in this protocol, and no numbers are claimed with regards to the accuracy, AUC, sensitivity, and specificity as a result of this work. The framework is designed to be used as a screening and decision-support pipeline to support, but not replace, the judgement of ophthalmologists. Keywords: Diabetic retinopathy; glaucoma; age-related macular degeneration; deep learning; convolutional neural network; transfer learning; medical imaging; ophthalmic screening.