A team led by Khon Kaen University applies an EfficientNetB7 convolutional neural network to color fundus photographs, classifying glaucoma severity according to the Hodapp-Parrish-Anderson criteria via transfer learning and fine-tuning. This approach offers accurate, single-image glaucoma screening in low-resource settings.

Key points

  • EfficientNetB7 CNN, pre-trained on ImageNet, classifies 2,940 fundus images into three glaucoma stages.
  • Transfer learning freezes 61% of layers and fine-tunes remaining layers for domain adaptation.
  • Model achieves overall accuracy 0.871 and AUCs of 0.988 (normal), 0.932 (mild-moderate), 0.963 (severe).

Why it matters: This AI-driven grading tool enhances early glaucoma detection and prioritizes severe cases, improving vision-loss prevention in resource-limited clinical settings.

Q&A

  • What is fundus photography?
  • What are Hodapp-Parrish-Anderson criteria?
  • How does transfer learning improve model performance?
  • Why use EfficientNetB7 specifically?
  • What do AUC and accuracy metrics indicate?
Copy link
Facebook X LinkedIn WhatsApp
Share post via...


Read full article

Ocular Aging and Glaucoma

Glaucoma is an age-related eye disease characterized by progressive optic nerve damage and visual field loss. It often develops slowly, with no symptoms until advanced stages, making early detection crucial. Risk factors include increased intraocular pressure, genetics, thin corneas, and vascular dysregulation in the optic nerve head. As people age, structural changes in the retina and optic disc make the eye more susceptible to damage from these factors.

Aging affects ocular blood flow, nerve fiber layers, and intraocular pressure regulation. Over time, cumulative stress on retinal ganglion cells leads to cell death, optic disc cupping, and corresponding visual field defects. Detecting these subtle changes early is essential for preventing irreversible vision loss and improving long-term quality of life for older adults.

Fundus Photography for Eye Health Screening

Fundus photography uses specialized cameras to capture high-resolution images of the retina, optic disc, and posterior pole. The procedure involves dilating the pupil with eye drops, positioning the patient at a retinal camera, and illuminating the eye with a flash. Fundus images reveal details like nerve fiber layer thickness, optic cup-to-disc ratio, and retinal blood vessels.

  • Non-invasive and quick: typically under one minute per eye.
  • Widely available in primary care and community clinics.
  • Enables teleophthalmology: images can be sent to specialists for remote evaluation.

Regular fundus photography screening allows clinicians to monitor changes over time, identify early signs of glaucoma, diabetic retinopathy, and macular degeneration, and intervene before significant vision loss occurs.

Deep Learning and Medical Imaging

Deep learning refers to neural networks with multiple layers that learn hierarchical feature representations directly from data. In medical imaging, convolutional neural networks (CNNs) excel at analyzing visual patterns such as textures, shapes, and edges. These models automatically detect disease markers without manual feature extraction.

  1. Training: Networks learn from labeled images (disease vs. healthy).
  2. Validation: Performance is tested on unseen images to tune hyperparameters.
  3. Inference: The final model predicts disease presence or severity on new images.

Deep learning enhances diagnostic speed and accuracy, reduces interobserver variability, and enables scalable screening programs.

EfficientNet and Transfer Learning

Transfer learning leverages CNNs pre-trained on large datasets like ImageNet. By freezing early layers that capture generic features and fine-tuning later layers on specific medical images, models adapt faster with less data. EfficientNet architectures optimize performance by balancing network depth, width, and input resolution. EfficientNetB7, a high-capacity variant, achieves strong accuracy on classification tasks with fewer parameters.

In glaucoma screening, transfer learning with EfficientNetB7 enables rapid development of models that distinguish normal, mild-moderate, and severe disease from fundus photographs, even in resource-limited settings.

Importance for Longevity and Vision Preservation

Maintaining good vision is vital for healthy aging and independent living. Early glaucoma detection prevents irreversible optic nerve damage and preserves quality of life. Integrating AI-based screening into routine check-ups empowers clinicians to identify high-risk patients earlier, initiate timely treatment, and reduce the burden of vision impairment in aging populations worldwide.

Machine learning technology in the classification of glaucoma severity using fundus photographs