Related Experiment Video
Updated: Feb 26, 2026
11:21
Integrated Photoacoustic Ophthalmoscopy and Spectral-domain Optical Coherence Tomography
Published on: January 15, 2013
12.0K
OpthaNet: Attention-Integrated Architecture for High-Precision Multi-Class Ophthalmic Image Classification.
Souhardo Rahman1, Md Nasif Safwan1, Mahamodul Hasan Mahadi1
1Department of Computer Science American International University-Bangladesh Dhaka Bangladesh.
Healthcare Technology Letters
|February 25, 2026
Summary
This study compared deep learning models for classifying eye diseases like cataracts, diabetic retinopathy, and glaucoma. Optimized models showed significant accuracy improvements, demonstrating AI
Area of Science:
- Ophthalmic diagnostics
- Artificial intelligence in healthcare
- Medical imaging analysis
Background:
- Deep learning models, including CNNs and transformers, are increasingly used in ophthalmic diagnostics.
- A direct comparison of these models for multi-class eye disease classification is limited.
- High-performing systems often require substantial computational resources, posing challenges for practical screening.
Purpose of the Study:
- To investigate and compare the efficacy of pre-trained deep learning models for multi-class classification of cataract, diabetic retinopathy, and glaucoma.
- To address practical bottlenecks in ophthalmic transfer learning, such as feature selectivity and overfitting with limited data.
- To evaluate tailored modifications for EfficientNetB3, MobileNetV2, and Vision Transformer models.
Main Methods:
- Evaluation of EfficientNetB3, MobileNetV2, and Vision Transformer models with specific customizations.
- Implementation of an attention-enhanced feature refinement module and OpthaHead classifier for EfficientNetB3 and MobileNetV2.
- Application of META customization to optimize the Vision Transformer model.
- Training and validation using fundus images for multi-class classification of eye diseases.
Main Results:
- Optimized EfficientNetB3 achieved 96.04% accuracy, a 10.84% improvement over baseline.
- Optimized MobileNetV2 showed an 11.26% improvement, balancing accuracy and computational efficiency.
- META-customized Vision Transformer performance increased by over 18%, indicating benefits of reduced complexity on limited medical data.
Conclusions:
- AI-driven classification demonstrates strong performance for detecting common eye diseases.
- Tailored model modifications can significantly enhance accuracy and efficiency in ophthalmic diagnostics.
- AI tools hold substantial potential for early eye disease detection, improving clinical decisions and patient outcomes.
