محمد الموسوي

This thesis presents an advanced multi-class, multi-label classification model for retinal disease detection using an architecture which includes a Vision Transformer (ViT) combined with Low-Rank Adaptation (LoRA). The proposed system is designed to overcome the limitations of traditional convolutional neural networks (CNNs) by leveraging the transformer’s self-attention mechanism and LoRA’s efficient fine-tuning capabilities, optimizing the model for retinal image classification. The model was eva‎luated using the ODIR-2019 dataset and achieved impressive results, including an accuracy of 0.97, an F1-score of 0.85 and ROC of 0.96, effectively detecting multiple retinal diseases such as diabetic retinopathy, glaucoma, and age-related macular degeneration. The use of LoRA not only minimized computational demands but also ensured generalization without overfitting, as demonstrated by the consistency of training and validation accuracy and error trends. These findings establish the ViT-LoRA model as a scalable, efficient, and robust solution for automated retinal disease detection, offering a significant improvement over existing CNN-based methods in both performance and adaptability, and paving the way for integration into real-world clinical applications.

ترانسفورمر بينايي , سازگاري با رتبه پايين , تشخيص بيماري شبكيه , رده‌بندي چند برچسبي چند رده , يادگيري عميق

Vision Transformer , Low-Rank Adaptation , Retinal Disease Detection , Multi-class Multi-label Classification , Deep Learning

Mohammed AlMusawi

Dr. Mohsen Soryani