Screening for Cataract.
From a Smartphone.
Over 4.1 million Indians are blind due to untreated cataract — a condition that is 92.9% treatable. Jilo Health's deep learning model brings specialist-grade cataract screening to the last mile using nothing but a mobile camera.
Cataract is India's leading cause of blindness.
And it is almost entirely preventable.
According to India's National Blindness & Visual Impairment Survey, cataract accounts for 66.2% of all blindness and 80.7% of severe visual impairment in the country. The survey documented 4.1 million blind and 29.2 million visually impaired individuals — of whom 92.9% had a treatable or preventable condition.
The healthcare infrastructure gap compounds the crisis. India has only 15 ophthalmologists per million population, and the optometrist-to-ophthalmologist ratio stands at 0.85 — far below the WHO recommendation of 3 optometrists per ophthalmologist.
Critically, only 30% of these limited resources are accessible in non-metro Bharat, where the burden of disease is highest and preventive care is nearly absent.
The cumulative economic toll is staggering — a $535 billion loss in Gross National Income due to avoidable blindness, a figure that has tripled in the past two decades.
Community Outreach Screening
- 01 Expensive fundus cameras (₹2L+ starting cost)
- 02 Logistically challenging to deploy in rural areas
- 03 AI models require specialist-grade fundus images
- 04 Significant infrastructure and trained operator dependency
- 05 Late detection · High treatment cost
Smartphone-Based AI Screening
- 01 Mobile camera — zero additional hardware cost
- 02 Works in any setting: camp, clinic, home
- 03 Deep learning model processes standard photos
- 04 Any frontline health worker can operate it
- 05 Instant result · Early referral · Better outcomes
The complete screening pipeline.
Deep learning + clinical validation.
Recent breakthroughs in deep learning have unlocked a new paradigm: processing mobile camera images at very high accuracy for medical screening tasks. Jilo Health's cataract detection model leverages this to replace expensive fundus cameras with the smartphone already in every health worker's pocket.
Capture
A frontline worker photographs the patient's eye using a standard smartphone. Our guided reticle overlay ensures correct alignment and optimal image quality.
AI Analysis
A compressed Vision Transformer (DINOv2) extracts deep features from the eye region. PCA reduces dimensions and an RBF SVM delivers the diagnostic classification.
Result + Referral
The model returns an instant cataract risk assessment. Positive cases are flagged for ophthalmologist referral; results are validated through Vision Function Questionnaires.
"The complete solution = Deep learning cataract detection + Vision Function Questionnaires to validate AI predictions."
Compressed QAT DINOv2 pipeline.
96%+ accuracy at 30.4 MB.
Training Dataset
Source: NHA website + open-source medical datasets. Train / Validation / Test split: 70 / 15 / 15.
Model Evaluation Results
| Class | Precision | Recall | F1-Score |
|---|---|---|---|
| Normal | 0.95 | 0.98 | 0.97 |
| Cataract | 0.98 | 0.94 | 0.96 |
| Macro Avg | 0.97 | 0.96 | 0.96 |
| Weighted Avg | 0.96 | 0.96 | 0.96 |
Structural Pruning
Dropped the final 4 layers of the standard 12-block DINOv2-small transformer, reducing to 15.0M parameters while retaining 99% of early edge and pattern sensitivities critical for lens opacity detection.
Quantization-Aware Training
MLP linear projections were fine-tuned with fake-quantization nodes. Post-training static INT8 conversion shrinks the model from ~90MB to 30.4MB — deployable on CPU-only environments.
PCA + RBF SVM Inference
High-dimensional 384D embedding vectors are compressed via PCA to 64 components to reduce noise. An RBF Support Vector Machine then classifies lens density into Normal or Cataract.
Vision Function Questionnaire (VFQ)
Beyond the scan: a validated 9-item patient-reported outcome measure (NEI-VFQ adapted) that captures the functional burden of cataract that an image alone cannot.
Highly complementary to neural network classifications, ensuring that patient referrals are based on both physiological lens opacity and actual daily-life visual impairment.
Why It Complements the AI Scan
Captures the functional visual disability that lens opacity images cannot. The combined AI + VFQ referral signal is far stronger for ASHA workers and community nurses.
What It Measures
9 questions on halos, glare, blurry/foggy vision, night mobility, recognising faces, reading small text, daily task performance, and quality of life impact. Scored 1 (no problem) to 5 (severe).
Scoring & Risk Bands
Total questionnaire score is normalized from 0 to 100. Risk bands: 0–30 Normal (self-monitor) · 30–50 Moderate (refer for clinical evaluation) · 50–100 Severe (urgent surgical referral).
Upload an eye image or use
your device camera.
Follow the alignment reticle overlays and capture or upload a clear, close-up image of the patient's eye. Results appear in seconds.
Upload Eye Scan
Drag & drop your eye scan image here, or select a file to upload.
Awaiting Diagnostics Scan
Acquire Macro Pupil Focus
Align the patient's iris under uniform, diffuse ambient lighting. Avoid bright specular glare on the cornea.
Feature Map Extraction
A compressed 8-layer Vision Transformer (DINOv2) processes the optical image to extract key structural parameters.
Support Vector Classification
Features are compressed via PCA to 64 components and classified by an RBF SVM to evaluate lens density.
Assess your daily visual challenges and symptoms.
Based on the National Eye Institute Visual Function Questionnaire (NEI-VFQ) ↗ — a peer-reviewed, validated patient-reported outcome instrument — adapted into a rapid 9-item screener for point-of-care use.
Symptom & Vision Burden Checklist
Cataract affects more than just visual sharpness—it changes how you perceive light, contrast, and fine details in daily life. This 9-question assessment evaluates your visual burden.
From pilot to certified Software as a Medical Device.
TRL 5
Technology Readiness Level
Prototype development
Model & interface completed
Clinical validation
In progress
CDSCO Class B SaMD
Certification filing