RetinaMD

Did you know that your eyes can reveal far more than what you see through them? They are not just windows to the soul, they are windows to your health. From diabetes to neurological disorders, to infections and inflammation, your eyes quietly hold the early signs of disease, waiting to be read. But here’s the problem: in most parts of the world, people don’t have access to specialists or expensive diagnostic equipment. Eye diseases often go unnoticed until it’s too late.

The Problem Every year, millions lose vision from conditions that could have been detected early: cataracts, uveitis, conjunctivitis. Yet, something as simple as a photograph could change that story. The Solution - Introducing RetinaMD At RetinaMD, we built a machine learning system that detects eye diseases using just a simple image of your eye, no lab, no hospital, just intelligent analysis. Our system doesn’t just tell you what’s wrong; it understands symptoms, predicts patterns, and guides you on what to do next.

How It Works To make RetinaMD both accurate and interpretable, we designed a two-stage training pipeline, inspired by how real doctors think. In the first stage, our model learns to recognize symptoms …. not diseases. We trained it on eye images and reported symptoms, excluding the final diagnosis. This forces the model to focus on visual cues: redness, irritation, cloudiness, instead of memorizing dataset-specific biases. This phase acts like a doctor’s first step: identifying what they see before concluding what it means. It also improves generalization and prevents overfitting. In the second stage, we combined the predicted symptoms with the original image to train a higher-level model - our Disease Predictor. This model uses a dual-branch CNN architecture: • The image branch captures visual features through convolutional layers. • The symptom branch processes 16 predicted symptom probabilities through dense layers. The two are then fused to classify the disease: Cataract, Conjunctivitis, Eyelid Disorders, Uveitis, or Normal. This layered reasoning allows the system to both explain and predict - you can see not just the disease, but also the symptoms that led to that conclusion. From a technical standpoint, our Symptom Predictor is built on MobileNetV2 for efficient transfer learning and outputs probabilities for 16 key symptoms like redness, pain, itching, tearing, and blurred vision. Inference time? Under 200 milliseconds per image, making it viable for real-time screening. By combining visual intelligence with symptom correlations, this two-stage design delivers speed, accuracy, and transparency: making RetinaMD not just a detection tool, but a step toward explainable medical AI.

The Vision But this is just the beginning. Imagine a world where your phone camera becomes your first health check. Where eye scans can detect not just diseases but even insulin levels, blood pressure patterns, or vision power adjustments. Imagine rural clinics, NGOs, and schools using RetinaMD to screen hundreds of people with a simple photo as well as hospitals integrating our API into their systems, giving instant feedback to ophthalmologists before a patient even steps into the room.

Our Learning While working on this project, we realized just how importantx it is to have a well-structured and balanced dataset. The challenges we faced taught us that having too little data limits what the model can learn, while too much unrefined data can introduce noise and reduce performance. Finding that balance was key to improving the overall reliability of our system.

The Impact With RetinaMD, we’re making early detection accessible, affordable, and intelligent. We’re not replacing doctors - we’re empowering them. We’re not just predicting disease - we’re preventing blindness.

Built With

• keras
• mobilenetv2
• numpy
• pandas
• python
• scikit-learn
• streamlit
• tensorflow