Explainable AI for Diabetic Retinopathy Screening in Rural India
“Rural India lacks enough eye doctors to check millions of diabetic patients for vision loss, leading to preventable blindness. We are building an AI tool using MATLAB that takes mobile eye camera photos, checks their quality, spots early signs of eye damage, and explains its decision to doctors in seconds.”
Deconstructed Problem Blueprint
WHAT TO BUILD?
An end-to-end MATLAB and Simulink-based retinal image analysis and telemedicine pipeline featuring automated quality check, multi-structure segmentation, severity classification, and a clinical explainability module.
WHO IS IT FOR?
Rural healthcare workers and primary health center (PHC) technicians capture the images, remote ophthalmologists review the validated cases, and diabetic patients across rural India are the primary beneficiaries preventing blindness.
WHY DOES IT EXIST?
There is a severe shortage of ophthalmologists in rural areas (~1 per 100,000 population), manual screening is impossible at scale, and existing AI systems are black boxes that clinicians do not trust, failing severely on low-quality portable field camera images.
INPUT → PROCESS → OUTPUT ARCHITECTURE PIPELINE
Raw retinal fundus images captured from portable hand-held fundus cameras, patient metadata (age, diabetes duration), and clinical grading guidelines.
Image quality assessment and CLAHE enhancement, retinal structure segmentation (vessels, optic disc, microaneurysms), International Clinical DR severity classification (Levels 0-4), Grad-CAM heatmap generation, and Simulink resource allocation simulation.
Automated PDF screening reports with lesion-level visual explanations, confidence scores, and referral recommendations delivered to doctors within 30 seconds, plus a district-level bandwidth and throughput optimization dashboard.
CONSTRAINTS
- Must run reliably on variable-quality portable fundus camera images from field conditions
- Must achieve >90% sensitivity and >85% specificity for referable DR (Level 2+)
- Must provide clinically interpretable Grad-CAM outputs and lesion-level evidence
- Must be built using MATLAB and Simulink toolboxes as mandated by MathWorks
- Ophthalmologist validation workflow must complete in under 30 seconds
JURY METRICS
- Sensitivity (>90%) and Specificity (>85%) for referable DR detection
- Microaneurysm and vessel segmentation accuracy (Dice coefficient / IoU)
- Image quality assessment accuracy (rejection rate of ungradeable images)
- Clinician usability score of the Grad-CAM explainability module
- Simulink pipeline processing throughput and bandwidth optimization metrics
CORE BOTTLENECK
Handling extreme variability and artifacts in portable fundus camera images captured by untrained field workers, while maintaining sub-pixel microaneurysm detection accuracy and true clinical trust via explainability.
Why You Should Choose This PS (Pros)
- Directly sponsored by MathWorks with explicit toolchain requirements (MATLAB, Simulink, Image Processing, Computer Vision, Deep Learning Toolboxes)
- High societal impact targeting preventable blindness in rural India matching national healthcare missions
- Public benchmark datasets (APTOS, IDRiD, Messidor-2) are readily accessible for model training and validation
- Clear, quantifiable metrics (>90% sensitivity, >85% specificity) make evaluation objective and winnable
Challenges & Risks to Watch (Cons)
- Strict enforcement of MATLAB/Simulink ecosystem might intimidate teams reliant purely on Python/PyTorch
- Sub-pixel microaneurysm detection requires rigorous preprocessing and fine-grained segmentation models
- Crowded healthcare track means differentiation requires exceptional explainability and a flawless Simulink simulation
- Handling class imbalance across DR severity levels (Level 0 vs Level 4) requires careful data augmentation
Recommended Tech Stack & Grand Finale Features
Killer Features to Impress the SIH Jury
View Official SIH 2026 Portal Raw Text▼
India has over 77 million diabetic adults
Design a MATLAB-based retinal image analysis pipeline for automated DR screening addressing real-world deployment challenges:
This problem demands clinical validation rigor, sub-pixel microaneurysm detection, and clinically meaningful explainability
A working prototype demonstrating: DR classification with >90% sensitivity and >85% specificity for referable DR;
explainable Grad-CAM outputs rated as clinically useful; a Simulink model optimizing screening resource allocation; and validation against published benchmarks showing the integrated pipeline outperforms any single technique approach.