Accuracy Efficient VLSI Architecture for Retinal Disease Detection using Deep Learning Technique

Abstract

This paper proposes accuracy‑efficient Very‑Large‑Scale Integration (VLSI) architecture for retinal disease detection using deep learning. We target high‑throughput, low‑latency inference on edge devices such as portable fundus cameras and ophthalmic screening kiosks where energy and memory budgets are constrained. The method combines (i) a compact, attention‑augmented convolutional backbone specialized for retinal lesions, (ii) mixed‑precision quantization with calibration‑aware retraining to retain diagnostic fidelity, and (iii) a memory‑centric dataflow that minimizes off‑chip transactions via tiling and on‑chip reuse. On public retinal image datasets, the proposed solution aims to achieve AUC ≥ 0.95 for diabetic retinopathy (DR) grading. The field of ophthalmology relies on digital image processing techniques, such as Optical Coherence Tomography (OCT), for diagnosing retinal diseases. However, manual interpretation of OCT images is time-consuming and prone to human error. This study developed a deep learning-based model to assist in diagnosing retinal pathologies from OCT images. VGG16 architecture was trained on a dataset of OCT images to classify four retinal conditions: choroidal neovascularization, diabetic macular edema, drusen, and normal. Rigorous evaluation, including cross-validation and independent testing, demonstrated the model’s ability to achieve a high accuracy and minimize loss.