Netra-AI: An AI-Powered Screening Tool to Prevent Diabetic Blindness
- The Problem: A Silent Epidemic in Rural India
Diabetic Retinopathy (DR) is a leading cause of preventable blindness among India's over 100 million diabetics. Early screening is critical, but the necessary equipment is expensive (over ₹12.5 Lakh) and concentrated in urban centers, leaving millions in rural areas unscreened and at risk of irreversible vision loss. Netra-AI aims to bridge this critical gap in public healthcare.
- Our Solution: Accessible AI for Early Detection
Netra-AI is an ultra-affordable (< ₹42,000 target cost), portable, AI-powered screening tool designed for use in primary health clinics and remote health camps. It empowers community health workers to conduct on-the-spot retinal screenings without needing an internet connection or a specialist's presence.
Key Features
-Instant Analysis: Provides a binary "Disease Present / No Disease" classification in seconds.
-Offline First: All AI processing happens directly on the edge device, ensuring functionality in the most remote areas.
-High Accuracy: The model is highly reliable, ensuring very few sick patients are missed.
-Ultra-Low Cost: Designed with low-cost hardware (Raspberry Pi target) to enable scalable deployment.
-Simple Interface: Can be operated by minimally trained personnel.
- The AI Model: A Two-Stage Training Strategy
To overcome the challenges of noisy and imbalanced medical data, we developed a robust training strategy.
Step A: Simplification
The initial complex 5-level grading task was simplified into a more stable binary classification problem: detecting the presence of DR versus its absence.
Step B: Two-Stage Transfer Learning
Foundation Training (Generalist Model): A powerful EfficientNetB0 model was first trained on the large, diverse APTOS 2019 dataset (3,600+ images). This built a strong "generalist" model capable of recognizing DR features from a wide variety of sources.
Demographic Specialization (Expert Model): The high-performing generalist model was then fine-tuned on the Indian-specific IDRiD dataset. This crucial step specializes the model's knowledge, enhancing its fairness and accuracy for the target Indian demographic.
- Final Model Performance
The final specialized model was evaluated on the unseen APTOS validation set, achieving outstanding results for a medical screening tool.
These metrics confirm that the model is both highly sensitive (low false negatives) and reliable (low false positives).