Accelerating Retinal Disease Detection Through Lightweight CNN Architectures on Edge Computing Devices

A Comparative Performance Analysis for OCT-Based Diagnosis

Abstract

This study presents a comprehensive evaluation of lightweight convolutional neural network (CNN) architectures for automated retinal disease detection using optical coherence tomography (OCT) images, with a specific focus on deployment feasibility for edge computing devices. Four distinct CNN architectures were systematically compared: ResNet50, EfficientNetB0, MobileNetV2, and a custom TinyCNN model, utilizing a balanced dataset of 32,064 training images and 968 test images across four disease categories (CNV, DME, DRUSEN, and NORMAL). The experimental results demonstrate that while ResNet50 achieved the highest test accuracy of 98.44%, the custom TinyCNN model delivered competitive performance at 97.29% accuracy with significantly reduced computational requirements (4.73 MB model size vs. 90.98 MB for ResNet50 and 4.20 seconds inference time vs. 14.53 seconds). MobileNetV2 emerged as an optimal balance between performance and efficiency, achieving 97.92% accuracy with a 9.24 MB model size and 9.05 seconds inference time. Notably, EfficientNetB0 exhibited training instability with poor generalization performance (24.48% test accuracy), highlighting the importance of architecture-specific optimization for medical imaging tasks. The findings provide critical insights for healthcare practitioners and developers seeking to implement real-time retinal disease screening systems on resource-constrained edge devices, demonstrating that lightweight architectures can maintain diagnostic accuracy while enabling practical deployment in clinical settings with limited computational resources.