Efficient Multi-Label Retinal Disease Classification with CLIP, LoRA, and Shadow Loss on the OIA-ODIR Dataset

Automated diagnosis of retinal diseases from fundus images is critical for early detection and intervention. However, challenges such as class imbalance and rare diseases like glaucoma and hypertension hinder model performance, especially under limited computational resources. This study proposes a resource-efficient deep learning framework for multi-label retinal disease classification using the OIA-ODIR dataset. We leverage the pre-trained CLIP model, fine-tuned with Low-Rank Adaptation (LoRA), which significantly reduces trainable parameters while preserving accuracy. To further enhance classification, we introduce Shadow Loss, which improves feature separation by pulling same-class samples closer and pushing different-class samples apart, complementing the standard Binary Cross-Entropy loss. Our experiments explore various LoRA configurations (rank, alpha), Shadow Loss margins, and baseline models such as ViT. Evaluated on the on-site test set, the model achieved a Final Score of ۰.۸۱۴۴, with a Kappa of ۰.۸۳۱۰, F۱-score of ۰.۸۵۳۰, and AUC of ۰.۷۵۹۰. The results highlight strong performance in detecting rare diseases and demonstrate competitiveness against state-of-the-art models like Fundus-DeepNet and Sivaz and Aykut [۳۰], despite being trained with only five epochs on a Kaggle T۴ GPU. This framework shows promise for retinal disease diagnosis in real-world, resource-constrained settings.

کلیدواژه ها:

Retinal Disease Classification ، CLIP ، LoRA ، Shadow Loss ، OIA-ODIR ، Multi-Label Classification