Hatvit-Caf: A Hierarchical Attention–Based Hybrid Vision Transformer For Multi-Class Ocular Disease Classification

Accurate automated classification of ocular diseases is essential for early diagnosis, appropriate clinical management, and prevention of vision loss. However, image-based decision support systems remain challenged by subtle inter-class similarities, heterogeneous disease presentations, and class imbalance, particularly in inflammatory ocular conditions.

Single tertiary referral eye centre

HATViT-CAF is proposed as an attention-driven hybrid deep learning framework for multi-class ocular disease classification. The model integrates hierarchical attention, cross-modal fusion, and disease-specific gating atop a pretrained EfficientNet-B0 backbone, yielding a 290-layer architecture with approximately 5.3 million parameters. Trained on 1,838 retinal images across five classes—cataract, conjunctivitis, eyelid disorders, uveitis, and healthy controls—and evaluated on 460 held-out images, the framework demonstrates competitive diagnostic performance as assessed by accuracy, precision, recall, and F1-score.

HATViT-CAF achieved 96.96% accuracy and a mean F1-score of 96.22% on the test set. Class-wise F1-scores reached 97.7% for cataract, 94.8% for conjunctivitis, 96.8% for eyelid disorders, 99.2% for normal eyes, and 92.6% for uveitis. Despite being the least represented class, uveitis yielded a recall of 97.8%, reflecting strong sensitivity for clinically challenging inflammatory cases. Confusion matrix analysis confirmed minimal misclassification across visually similar disease categories.

The proposed HATViT-CAF model provides robust and reliable multi-class ocular disease classification despite class imbalance. Its attention-driven, disease-aware design and moderate model complexity support its potential use as a computer-aided decision support tool for ophthalmic screening and clinical triage in real-world settings.