Ensemble-Based Augmentation AI Framework for Enhanced Glaucoma Detection Using Deep Learning Models
Zahra Fathollahi1 *
- Novin Pars emerging intelligent health technologies, Shiraz, Iran.
Abstract: Glaucoma is a leading cause of irreversible blindness, particularly in adults over 40. Early detection is crucial to prevent vision loss; however, manual screening methods are time-intensive and require trained specialists. Automated diagnostic systems powered by artificial intelligence (AI) offer a promising alternative. This study introduces an AI-based system for glaucoma detection using fundus images to improve diagnostic accuracy and efficiency.
Methods: The system incorporates advanced preprocessing and ensemble learning techniques. Preprocessing steps, including contrast enhancement, brightness balancing, and noise reduction, were applied to improve image quality. The optic nerve head (ONH) region, comprising the optic cup and disc, was segmented and cropped to highlight critical features such as the cup-to-disc ratio (CDR), retinal nerve fiber layer (RNFL) thinning, and optic disc notching, key indicators of glaucoma. An ensemble of three deep learning models with complementary strengths was used.
The models classified fundus images as No Referable Glaucoma (NRG) or Referable Glaucoma (RG), with performance assessed using accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC).
Results: The system demonstrated strong performance in detecting glaucoma:
Training Datasets
• SMDG Dataset (12,316 images: 7,549 NRG, 4,767 RG): Accuracy of 91.3%, sensitivity of 87.2%, specificity of 93.4%, and AUROC of 0.97.
• JUSTRAIG Dataset (7,895 images: 4,750 NRG, 3,145 RG): Accuracy of 90.9%, sensitivity of 87.1%, specificity of 92.8%, and AUROC of 0.96.
External Validation Datasets
• EyePACS Light (8,000 images: 4,000 NRG, 4,000 RG): Accuracy of 86.3%, sensitivity of 86.2%, and specificity of 86.5%.
• GAMMA (66 images: 43 NRG, 23 RG): Accuracy of 93.9%, sensitivity of 95.6%, and specificity of 93.0%.
Conclusion: This AI-based system effectively identifies glaucomatous changes in fundus images, demonstrating high sensitivity and specificity. Automating the screening process reduces the workload for ophthalmologists and enables early diagnosis, particularly in resource-limited settings. Its robust performance across diverse datasets underscores its potential for clinical integration, helping prevent vision loss and improve outcomes.
