Multi-Label Ocular Disease Recognition Using Vision Transformers on the ODIR-5K Dataset
Document Type
Conference Proceeding
Publication Date
2026
Abstract
Ocular diseases are a leading cause of preventable blindness worldwide, affecting over 2.2 billion people. This paper proposes a deep learning-based model for multi-label ocular disease recognition using the ODIR-5K fundus image dataset, which covers 6,392 patients with both left and right eye fundus images across eight disease categories. The proposed model follows three main phases: (1) data preprocessing including random oversampling and comprehensive data augmentation; (2) a fine-tuned Vision Transformer ViT-B/16 with a custom residual classification head trained using label smoothing, warmup scheduling, and gradient accumulation; and (3) Grad-CAM visual interpretability analysis to validate clinical relevance. The overall proposed model achieved an accuracy of 90.45%, F1-Score of 93.83%, and area under the curve of 96.74%. The results as well revealed that the proposed model outperform all CNN-based methods evaluated on the same dataset.
Recommended Citation
El-Azm, A. A., & Sayed, G. I. (2026, July). Multi-Label Ocular Disease Recognition Using Vision Transformers on the ODIR-5K Dataset. In 2026 Intelligent Methods, Systems, and Applications (IMSA) (pp. 324-329). IEEE. https://doi.org/10.1109/IMSA70415.2026.11700329