Deep Learning for Diabetic Retinopathy Risk From Retinal Images

ISEF Category: Biomedical and Health Sciences

Subcategory: Pathophysiology  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: Full Year

The Hook

A tiny change in a retinal photo can hint at eye damage long before vision gets worse. One small bulge in a vessel may matter more than the whole image. Your project can test whether a self-supervised model spots those early warning signs better than a standard classifier. That matters because earlier risk detection can change who gets follow-up care.

What Is It?

Diabetic retinopathy is damage to the blood vessels in the retina, the light-sensing layer at the back of the eye. In retinal-fundus photos, one early clue is a microaneurysm, which is a tiny bulge in a vessel wall. Think of the retina like a street map. When a few roads start to swell or leak, the pattern tells you something is changing even before the eye looks obviously unhealthy.

A self-supervised foundation model learns general image patterns from many unlabeled photos before it gets fine-tuned on a smaller labeled set. In this topic, you would test whether that pretraining helps the model rank progression risk from subtle clues, like microaneurysm density, instead of only saying healthy or diseased. The main question is not just whether the model can spot disease, but whether it can tell which eyes look more likely to worsen.

Why This Is a Good Topic

This is a strong science fair topic because you can test a clear question with public data, repeatable metrics, and no wet lab work. It connects to a real problem, catching eye disease early, and it lets you learn image preprocessing, model training, and evaluation. You can also compare datasets, model types, and labeling choices, which gives you room for original analysis.

Research Questions

• How does self-supervised pretraining affect AUC for predicting diabetic retinopathy progression risk from retinal fundus images?
• What is the effect of adding microaneurysm density as an auxiliary target on risk ranking performance?
• Does a model trained on EyePACS transfer better to Messidor than a model trained from scratch?
• To what extent does class imbalance correction improve recall for high-risk cases?
• Which image preprocessing choice gives the most stable risk score across datasets?
• How does image resolution change sensitivity to early microaneurysm cues?

Basic Materials

• Laptop or desktop with at least 16 GB RAM and internet access.
• Free Google Colab account or school GPU access.
• Python 3.10 or later.
• Jupyter Notebook or VS Code.
• Access to EyePACS and Messidor public dataset files or their published subsets.
• Spreadsheet or lab notebook for tracking splits, metrics, and model runs.

Advanced Materials

• University GPU workstation or shared compute server.
• Python environment with CUDA-capable PyTorch.
• Secure storage for large image files and checkpoints.
• Annotation tool for lesion review or image quality checks.
• Access to an ophthalmology mentor or clinician for error review.
• Statistical analysis scripts for calibration, confidence intervals, and ablation tests.

Software & Tools

• Python: Runs the data cleaning, training, and evaluation code for the project.
• PyTorch: Trains the image model and supports transfer learning and fine-tuning.
• Jupyter Notebook: Keeps experiments, plots, and notes in one place.
• Google Colab: Gives free GPU access for smaller training runs.
• scikit-learn: Calculates AUC, confusion matrices, and calibration metrics.
• ImageJ: Lets you inspect retinal images and compare visible lesion patterns.

Experiment Steps

1. Define the exact prediction target, such as severity class, progression risk, or microaneurysm density.
2. Choose one baseline model and one self-supervised model so your comparison stays fair.
3. Plan a split strategy that keeps each patient in only one dataset split and tests EyePACS and Messidor separately.
4. Decide which metrics matter most, such as AUC, sensitivity for high-risk cases, and calibration.
5. Design ablation tests for pretraining, image resolution, and lesion-focused cropping.
6. Set up an error review plan for false negatives, false positives, and low-quality images.

Common Pitfalls

• Mixing images from the same patient across train and test splits, which makes performance look better than it is.
• Treating dataset labels as if they mean the same thing in EyePACS and Messidor, which can break cross-dataset comparisons.
• Letting borders, camera marks, or other image artifacts drive the prediction instead of retinal tissue.
• Reporting only accuracy, which can hide weak performance on the high-risk cases you care about most.
• Comparing models with different crops or resolutions without controlling for those changes, which muddies the result.

What Makes This Competitive

A strong version of this project would not stop at one model score. It would compare self-supervised and fully supervised backbones, test on a second dataset, and report calibration, not just accuracy. Better entries also explain errors with example images and show that the model still works when the dataset changes. That kind of analysis makes the work feel like real research, not just a training run.

Project Variations

• Train on EyePACS and test on Messidor to study dataset shift in retinal risk prediction.
• Swap the target from severity class to microaneurysm density and compare which label gives cleaner risk ranking.
• Compare a self-supervised backbone with a plain supervised CNN on the same images and report calibration, not just accuracy.

Learn More

• PubMed: Search review articles on diabetic retinopathy progression, retinal fundus imaging, and self-supervised vision models.
• National Eye Institute: Read the NIH overview of diabetic retinopathy and basic retina anatomy.
• EyePACS original paper: Find the dataset description and grading details through PubMed or the journal site.
• Messidor original paper: Read the dataset release paper and note the image quality and label setup through PubMed or the journal site.
• Ophthalmology: Search recent methods and review papers on diabetic retinopathy detection through PubMed or the journal site.
• Papers with Code: Compare published results for diabetic retinopathy detection and retinal self-supervised learning.