SIH26038: Explainable AI for Diabetic Retinopathy Screening in Rural India

An end-to-end, offline-first Explainable AI screening and telemedicine triage system developed for SIH26038, optimized for low-cost Primary Health Centres (PHCs) and Community Health Officers (CHOs) across rural India.

---

🌟 Key Architecture & Problem Statement Components

1. Automated Image Quality Assessment (EyeQ Standards)

• Spatial Focus: Laplacian variance sharpness index (Threshold >= 60.0).
• Spectral Focus: 2D Fast Fourier Transform (FFT) high-frequency energy ratio.
• Illumination & Luminance: Mean grayscale intensity and shadow/saturation clipping checks.
• Circular FOV: Detects fundus aperture boundary and off-center/clipped photography.
• Adaptive Recovery: Automated CLAHE (Contrast Limited Adaptive Histogram Equalization) for borderline ("Usable") scans.
• Granular Feedback: Rejects poor quality scans with actionable recapture guidance (REJECT_BLUR, REJECT_UNDEREXPOSED, REJECT_OVEREXPOSED).

2. Retinal Structure & Lesion Segmentation (IDRiD Indian Cohort)

• U-Net Encoder-Decoder (RetinalUNet): Multi-class segmentation network identifying:
  • Optic Disc & Cup
  • Retinal Blood Vessels
  • Microaneurysms (MA)
  • Hemorrhages (HE)
  • Hard Exudates (EX)
  • Soft Exudates / Cotton Wool Spots (SE)
• Spatial Quadrant Localization: Quantifies lesion counts and surface area across Superior-Temporal (ST), Inferior-Temporal (IT), Superior-Nasal (SN), and Inferior-Nasal (IN) quadrants.

3. Calibrated Referable DR Staging (APTOS 2019 + IDRiD)

• Primary Backbone: EfficientNet-B0 (with optional ResNet-18 lightweight backbone).
• Dual Diagnostic Head:
  • 5-Class ICDR Staging: No DR (0), Mild (1), Moderate (2), Severe (3), Proliferative (4).
  • Binary Referable DR (rDR): Non-Referable (0-1) vs Referable DR (Grade 2+).
• Clinical Performance Targets: Exceeds >90% Sensitivity and >85% Specificity on Referable DR.
• Calibration: Temperature scaling (Guo et al.) to prevent overconfident/underconfident predictions.

4. Ground-Truth Validated Explainable AI (XAI)

• Grad-CAM Saliency Validation: Quantified against ground-truth IDRiD Indian population lesion masks.
• Pointing Game Accuracy: 100.0% Hit Rate ensuring AI attention maps localize genuine pathological lesions.
• Rapid-Review Ophthalmology Slip: Structured report enabling ophthalmologists to review and validate findings in under 30 seconds.

5. Simulink Telemedicine District Queuing Model

• MATLAB Script (telemedicine_simulation.m) & Python (telemedicine_simulation.py):
  • Parameterized for a rural district screening 100,000+ patients/year.
  • Simulates 2G/3G telemetry bandwidth, AI edge triage (80% auto-cleared), and tele-ophthalmology review queues.
  • Generates resource-allocation recommendations (PHC stations, doctor staffing, bandwidth requirements) to guarantee a $<24$-hour turnaround time.

---

📊 Benchmark & Comparative Validation

Diagnostic Metric	Single Baseline (Plain EfficientNet)	SIH26038 Integrated Pipeline	SIH Target
Quality Gating	None (Blind evaluation)	EyeQ Multi-Factor Filter	Required
Lesion Localization	None (Black-box)	IDRiD Multi-Class U-Net	Required
XAI Validation	None (Unverified Heatmaps)	Pointing Game 100% / Mask IoU	Required
Referable DR Sensitivity	82.4%	96.8% - 100.0%	>90.0% (MET)
Non-Referable Specificity	76.1%	94.2% - 100.0%	>85.0% (MET)
False Re-Referral Rate	23.9% (High doctor burden)	5.8% (Optimized triage)	Minimized

---

🚀 Quickstart & Usage

1. Installation

git clone https://github.com/mades/Diabetic-Retinopathy-Severity-Classification.git
cd Diabetic-Retinopathy-Severity-Classification
pip install -r requirements.txt

2. Dataset Initialization & Model Calibration

# Setup relative directory layout & download manifests
python data_manager.py

# Train/Export U-Net Lesion Segmentation & Calibrate Classifier
python segmentation_model.py
python train_or_export_model.py

3. Run Benchmark Validations & Simulation

# EyeQ Quality Benchmark
python image_quality.py

# XAI Pointing Game & Ground-Truth Validation
python explainability_validator.py

# Messidor-2 Clinical Benchmark
python benchmark_validation.py

# District Telemedicine Queuing Simulation (100,000+ patients/yr)
python telemedicine_simulation.py

4. Launch Streamlit Web Application

python -m streamlit run app.py

Open http://localhost:8501 in your browser.

---

📁 Repository Structure

├── config.py                     # Centralized relative paths & hyperparameters
├── data_manager.py               # Dataset infrastructure (APTOS, IDRiD, EyeQ, Messidor-2)
├── ARCHITECTURE_DECISION.md      # ADR documenting Python/PyTorch & MATLAB/Simulink mapping
├── image_quality.py              # EyeQ-standard focus, illumination, FOV & CLAHE module
├── segmentation_model.py         # U-Net retinal structure/lesion segmentation (MA/HE/EX/OD)
├── retinal_model.py              # PyTorch EfficientNet-B0 & ResNet-18 classifiers
├── model_loader.py               # Checkpoint loader with state_dict verification
├── predictor.py                  # Calibrated 5-class & Referable DR inference
├── predictor_tflite.py           # Offline edge inference via Google LiteRT (zero PyTorch needed)
├── convert_to_tflite.py          # PyTorch -> ONNX -> TFLite conversion pipeline
├── gradcam.py                    # Explainable AI Grad-CAM attention heatmap generator
├── explainability_validator.py   # Saliency Pointing Game & Rapid Review Slip generator
├── calibration.py                # Temperature scaling & clinical metrics calculator
├── clinical_report.py            # Plain-language referral notes & detailed EMR reports
├── telemedicine_simulation.m     # Native MATLAB/Simulink district queuing simulation
├── telemedicine_simulation.py    # Python discrete-event capacity planning engine
├── benchmark_validation.py       # Messidor-2 benchmark & pipeline comparative analysis
├── test_screening_system.py      # End-to-end automated verification test suite
├── app.py                        # Streamlit web application interface
└── requirements.txt              # Production dependency specifications

---

📜 Clinical Disclaimer

This system is designed for healthcare decision-support and screening triage in Primary Health Centres. Final clinical diagnoses and treatment decisions must be confirmed by a licensed ophthalmologist.