VisionOcular: Transfer Learning with Self-Attention for Multi-Class Ocular Diagnosis
Students & Supervisors
Student Authors
Supervisors
Abstract
The growing integration of artificial intelligence with ophthalmic imaging has opened new possibilities for automating the detection and classification of eye diseases. However, many prior studies have been constrained by limited datasets and a narrow focus on a small number of disease categories. Addressing this gap, our study leverages a recently released dataset collected from real-world clinical hospitals, comprising 10 distinct eye disease classes to develop a more comprehensive multiclass classification framework. Our proposed approach uses a pre-trained deep learning model, enhanced with an attention mechanism to more effectively capture discriminative features and combined with a custom-designed classifier to enhance overall diagnostic performance. Among the models evaluated, our proposed EfficientNetV2S-CA architecture achieved an accuracy of 90.87%, representing a 5.54% improvement over the baseline model. These findings highlight the effectiveness of our approach in setting a solid benchmark for multiclass eye disease classification, paving the way for earlier diagnosis and more dependable clinical outcomes.
Keywords
Publication Details
- Type of Publication:
- Conference Name: 2025 IEEE International Conference on Quantum Photonics, Artificial Intelligence, and Networking
- Date of Conference: 31/07/2025 - 31/07/2025
- Venue: Bangladesh Army University of Science and Technology (BAUST), Rangpur
- Organizer: Bangladesh Army University of Science and Technology (BAUST)