Retina Reach

Inspiration

Diabetic Retinopathy (DR) can progressively damage vision, yet many people in rural and underserved communities face a major barrier: access to timely eye-specialist screening.

We were inspired by a simple question: What if the first step toward detecting a potentially sight-threatening condition could happen closer to the patient, even where an ophthalmologist is not available?

This led us to build Retina Reach — an AI-assisted retinal screening and referral system designed for resource-constrained healthcare settings.

Our goal is not to replace ophthalmologists. Instead, Retina Reach acts as a screening and triage layer: a trained health worker can capture a retinal image, the AI estimates the level of diabetic-retinopathy risk, and patients who may require specialist attention can be prioritized for referral to an ophthalmologist.

This approach addresses an important access bottleneck: a limited number of specialists should be able to spend more of their time evaluating patients who actually require specialist assessment, rather than manually screening every patient.

AI-assisted screening — not a diagnosis. Final diagnosis and treatment decisions remain with a qualified ophthalmologist.

What We Learned

Building Retina Reach taught us that healthcare AI is not simply a machine-learning problem.

A model can classify an image, but a useful healthcare system must also consider:

• Image quality and unreliable inputs
• Model uncertainty and false negatives
• Human in the loop clinical review
• Referral workflows
• Low connectivity environments
• Usability for minimally trained health workers
• Clear communication of AI limitations
• The difference between screening, diagnosis, and treatment

We also learned that the most meaningful innovation does not necessarily come from inventing a completely new technology . It can come from designing an existing technology around a real world healthcare constraint.

How We Built It

Retina Reach uses a computer vision pipeline for retinal fundus images.

Screening Pipeline

1. Image Capture

   • A retinal fundus image is captured or uploaded by a health worker.

2. Image Quality Check

   • The system checks whether the image is sufficiently usable for AI screening.
   • Poor quality or uncertain cases can be flagged for retaking of fundus images.

3. AI-Based Screening

   • A pretrained convolutional neural network will be fed on labeled retinal fundus images.
   • The model estimates the diabetic-retinopathy severity category.

4. Risk-Based Triage

   • The AI result is converted into a simple screening/referral recommendation rather than presented as a diagnosis.
   • Higher-risk or uncertain cases are prioritized for ophthalmologist review.

5. Ophthalmologist Review

   • Referred cases reach an ophthalmologist at a district hospital, larger healthcare facility, or through tele-ophthalmology where available.
   • The ophthalmologist remains the final clinical decision-maker.

6. Follow-Up

   • The system can maintain screening and referral records so that patients are less likely to be lost between screening and specialist evaluation.

Technology

The prototype is designed around a modular architecture:

• AI/ML: Python with TensorFlow/Keras or PyTorch
• Model: Transfer learning using an image-classification architecture such as EfficientNet
• Backend: FastAPI/Flask
• Frontend: Web-based health-worker interface
• Training data: Publicly available labeled retinal fundus datasets
• Deployment concept: Low-bandwidth/offline-capable inference using an optimized model

The interface is intentionally simple, with a clear visual risk indication and minimal clinical jargon so that it can be used in primary-care environments.

We considered the realities of rural healthcare: limited connectivity, limited specialist availability, varying levels of digital literacy, and the need for affordable imaging hardware. Hence we will create a platform that will addresses all this problem.

Our Vision

Retina Reach is designed around a human + AI model of eyecare:

Health Worker → AI Screening → Risk-Based Referral → Ophthalmologist → Final Clinical Decision

The long-term vision is to make specialist-supported retinal screening without removing the specialist from the loop.

Retina Reach does not aim to replace doctors.

It aims to help the right patient reach the right doctor at the right time.

Built With

• ai
• fastapi
• healthcare
• imaging
• keras
• medical
• ml
• python
• tensorflow
• vision