Multi-Class Classification Using Optimized Hybrid Deep Neural Network for Retinopathy Detection

Abstract

Timely detection of retinal pathology is among the most demanding challenges in ophthalmic practice. Lesion signatures in early-stage disease are extremely subtle, and the supply of trained graders falls well short of global screening demand. We introduce HybridDL, a four-class automated classification system that pairs a carefully tuned Contrast Limited Adaptive Histogram Equalization (CLAHE) preprocessing stage with a ResNet50 convolutional feature extractor and a 200-tree Random Forest (RF) decision layer. Preprocessing is confined t o the luminance channel of the CIELAB colour space so that blood-vessel morphology, microaneurysm contrast, and haemorrhage boundaries are sharpened without distorting the chromatic cues that ophthalmologists rely upon during manual review. The four target categories — Diabetic Retinopathy (DR), Retinopathy of Prematurity (ROP), Hypertensive Retinopathy (HTR), and Central Serous Retinopathy (CSR) — share superficially s imilar fundus appearances yet demand entirely different clinical responses. Evaluated on 112 balanced test images, HybridDL records 91 % accuracy, 89 % precision, 92 % recall, and 90 % F1-score, against 74 %, 70 %, 72 %, and 73 % for an unenhanced CNN baseline. A memoisation layer in the Streamlit delivery interface reduces per-image inference overhead for repeated queries by a factor of 40, bringing gallery-scale screening within reach of commodity clinic hardware.