Comparative Analysis of Deep Learning Architectures for Retina Disease Detection

Retinal diseases such as diabetic retinopathy (DR), agerelated macular degeneration (AMD), and glaucoma are some of the major causes of preventable blindness in the world. Screening at the earliest stage is very important but manual processes have a limitation of subjectivity, time consuming, and inaccessibility. The current paper examines the use and the performance of four state-of-the-art deep learning frameworks, including Convolutional Neural Networks (CNNs), You Only Look Once version 8 nano (YOLOv8n), Vision Transformers (ViT), and Swin Transformers, to detect retinal pathologies with the help of the Retinal Fundus Multi-disease Image Dataset (RFMiD). We assess these models based on accuracy, precision, recall, F1- score and computational endeavors. The CNNs are used as a baseline, and the goal of the YOLOv8n is to perform localization of lesions in real-time, and the ViT and Swin Transformers apply attention processes to analyze the global context. Based on our findings, the Swin Transformer has the best accuracy (95.43), and balanced F1-score (0.9540), indicating that it is capable of detecting local lesions and global retinal structures. ViT and CNNs have comparable accuracy (around 90%), whereas YOLOv8n is efficient in inference and can be used in resource-constrained settings even though its accuracy is lower $(89.06 \%)$. This paper demonstrates the promise of transformer-based models and specifically Swin to deliver highquality AI-assisted retinal diagnostics to facilitate the availability of efficient and readily accessible screening to fight avoidable blindness.