-green.svg){kind=icon}
Problem Statement ID: SIH26038
Title: Explainable AI for Diabetic Retinopathy Screening in Rural India
Subsystem: AI Engine — Classical Computer Vision & Explainable Machine Learning
Diabetic Retinopathy (DR) is a primary cause of preventable blindness worldwide, disproportionately impacting rural and semi-urban populations where access to retinal specialists and tertiary ophthalmology clinics is limited.
Retino-AI is designed for rural point-of-care screening:
- No Costly Cloud GPUs or Black-Box Deep Networks: Built on robust, deterministic classical computer vision and ensemble machine learning that runs locally on modest field clinic hardware.
- Strict Quality Gatekeeping: Automatically inspects images for focus, underexposure, overexposure, and field-of-view (FOV) validity, instantly rejecting ungradable photographs before classification.
- Mask-Free Clinical Inference: Does not require ground-truth lesion annotations or manual vessel masks. Newly captured patient fundus photographs are evaluated directly from raw pixels.
- Explainable Clinical Biomarkers: Translates fundus imagery into 12 measurable color and texture biomarkers (photometric distribution, green-channel microvascular contrast, and Gray-Level Co-occurrence Matrix texture descriptors).
Raw Retinal Fundus Photograph
│
▼
┌──────────────────────────────────────────────────────────┐
│ 1. Quality Assessment Gatekeeper │
│ - Focus / Sharpness Metric (Variance of Laplacian) │
│ - Exposure Check (Mean luminance inside retinal FOV) │
│ - Retinal Coverage & Circularity Detection │
└────────────────────────────┬─────────────────────────────┘
│
Is image gradable for diagnosis?
├── NO ──► [UNGRADABLE] Halted with Clinical Advisory
└── YES
│
▼
┌──────────────────────────────────────────────────────────┐
│ 2. Fundus Preprocessing │
│ - Retinal Field of View (FOV) auto-crop & margin padding│
│ - Standardization to 512 x 512 resolution │
│ - Green-Channel Isolation (peak hemoglobin contrast) │
│ - Illumination Gaussian normalization & CLAHE │
└────────────────────────────┬─────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────┐
│ 3. Mask-Free Feature Extraction │
│ - 8 Photometric Descriptors (Mean & Std RGB, Contrast) │
│ - 4 Haralick Texture Metrics (GLCM Contrast, Energy, │
│ Correlation, Homogeneity across 4 directions) │
└────────────────────────────┬─────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────┐
│ 4. Random Forest DR Classifier │
│ - 100 Bagged Decision Trees (Min Leaf Size = 5) │
│ - Balanced subsample weighting across ICDR Grades 0-4 │
│ - Multi-class posterior probability distribution (1x5) │
└────────────────────────────┬─────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────┐
│ 5. Diagnostic & Referral Output │
│ - Predicted ICDR Severity Grade (0 to 4) │
│ - Screening Confidence Score (0.0 to 1.0) │
│ - Referable DR Flag: TRUE if Grade >= 2 │
│ - Structured JSON Contract & MATLAB Diagnostic Struct │
└──────────────────────────────────────────────────────────┘
Because incoming patient photographs in rural camps have zero prior manual segmentations, the classifier operates exclusively on directly-computable, mask-free biomarkers:
| # | Feature Name | Clinical Rationale | Extraction Domain |
|---|---|---|---|
| 1 | mean_red |
Overall retinal tissue reflectance and background illumination | Retinal FOV |
| 2 | mean_green |
Blood vessel and retinal hemorrhage absorption | Retinal FOV |
| 3 | mean_blue |
Optical media clarity / cataract haze scatter | Retinal FOV |
| 4 | std_red |
Spatial intensity dispersion across choroidal background | Retinal FOV |
| 5 | std_green |
Microvascular and hemorrhagic contrast variance | Retinal FOV |
| 6 | std_blue |
Blue channel variance across retinal surface | Retinal FOV |
| 7 | green_mean |
Dedicated green-channel mean luminance | Retinal FOV |
| 8 | green_contrast |
Dynamic intensity spread of the green working channel | Retinal FOV |
| 9 | glcm_contrast |
Local intensity variations in microvascular tissue | Green Channel GLCM (16 levels) |
| 10 | glcm_correlation |
Linear pixel-pair dependency across neighboring structures | Green Channel GLCM (4 directions) |
| 11 | glcm_energy |
Uniformity and order of the retinal textural pattern | Green Channel GLCM |
| 12 | glcm_homogeneity |
Closeness of GLCM element distribution to diagonal | Green Channel GLCM |
The Random Forest model was trained and evaluated on 4,075 clinical retinal images with a stratified 80% training / 20% held-out test split (fixed seed 42 for exact reproducibility).
| Evaluation Metric | Value | Clinical Significance |
|---|---|---|
| 5-Class DR Accuracy | 66.38% | Multi-class distinction across all 5 stages (chance = 20.0%) |
Referable DR Accuracy (Grade >= 2) |
82.45% | Correct determination of whether patient needs referral |
| Referable DR Sensitivity | 76.50% | Fraction of diseased eyes (Grade >= 2) correctly referred (267/349) |
| Referable DR Specificity | 86.91% | Fraction of healthy/mild eyes cleared without unnecessary hospital burden (405/466) |
| Healthy Eye (Grade 0) Identification | 92.01% | 357 out of 388 normal eyes accurately identified |
PREDICTED DR GRADE
0 (No DR) 1 (Mild) 2 (Mod) 3 (Severe) 4 (PDR) Total Sensitivity
Actual 0: 357 7 19 3 2 388 92.01%
Actual 1: 9 32 29 5 3 78 41.03%
Actual 2: 23 32 128 22 22 227 56.39%
Actual 3: 5 3 20 15 10 53 28.30%
Actual 4: 7 12 34 7 9 69 13.04%
Total Pred: 401 86 230 52 46 815
- True Positives (TP): 267
- False Negatives (FN): 82
- True Negatives (TN): 405
- False Positives (FP): 61
-
Negative Predictive Value (NPV): 83.16% (
$\frac{405}{405 + 82}$ ) — Very high reassurance that a negative screen is safe.
To ensure clinical integrity, datasets are strictly separated by functional roles:
| Dataset | Sample Count | Primary Role in Retino-AI |
|---|---|---|
| APTOS 2019 Blindness Detection | 3,662 | Primary 5-class DR severity classification data |
| IDRiD Disease Grading | 413 | Supplementary 5-class DR severity classification data |
| DRIVE | 40 | Retinal vascular tree ground-truth validation (NOT used for DR grading) |
| Messidor-2 | — | Skipped (not required for prototype) |
Retino-AI/
├── Backend/
│ ├── ai-engine/
│ │ ├── config/ # Pipeline & training configuration
│ │ │ ├── training_config.m # Hyperparameters, thresholds, feature settings
│ │ │ ├── dataset_config.m # Dataset paths and loaders config
│ │ │ └── config.json # Interoperable JSON configuration
│ │ │
│ │ ├── matlab/
│ │ │ ├── preprocessing/ # Stage 1: Preprocessing & FOV isolation
│ │ │ │ ├── load_fundus.m
│ │ │ │ ├── crop_fov.m
│ │ │ │ ├── resize_fundus.m
│ │ │ │ ├── extract_green_channel.m
│ │ │ │ ├── normalize_illumination.m
│ │ │ │ ├── enhance_contrast.m
│ │ │ │ └── preprocess_fundus.m
│ │ │ │
│ │ │ ├── quality_assessment/ # Stage 2: Automated Quality Gatekeeper
│ │ │ │ ├── check_blur.m # Focus / sharpness evaluation
│ │ │ │ ├── check_brightness.m # Exposure & illumination evaluation
│ │ │ │ ├── check_fov.m # Circularity & retinal coverage
│ │ │ │ └── assess_image_quality.m
│ │ │ │
│ │ │ ├── feature_extraction/ # Stage 3: Mask-Free Clinical Biomarkers
│ │ │ │ ├── extract_color_features.m
│ │ │ │ ├── extract_texture_features.m # GLCM Haralick texture descriptors
│ │ │ │ ├── extract_vessel_features.m # Vascular density metrics
│ │ │ │ ├── extract_lesion_features.m # Lesion evidence analysis
│ │ │ │ └── extract_features.m
│ │ │ │
│ │ │ ├── models/ # Stage 4: Model Training & Inference
│ │ │ │ ├── train/
│ │ │ │ │ ├── train_dr_model.m # Full training pipeline
│ │ │ │ │ └── train_dr_model_runner.py
│ │ │ │ ├── inference/
│ │ │ │ │ └── predict_dr_grade.m
│ │ │ │ └── saved/
│ │ │ │ └── dr_random_forest.mat # Trained model artifact
│ │ │ │
│ │ │ ├── inference/ # Master Single-Image Inference
│ │ │ │ ├── predict_dr.m # Master predict_dr() function
│ │ │ │ └── analyze_fundus.m # Complete explainability pipeline
│ │ │ │
│ │ │ ├── evaluation/ # Metrics, confusion matrix, sensitivity/specificity
│ │ │ ├── setup_paths.m # Adds folders to MATLAB path
│ │ │ └── startup.m
│ │ │
│ │ ├── models/saved/ # Saved Model Artifacts & Metadata
│ │ │ ├── dr_random_forest.mat # Native MATLAB tree ensemble (< 500 KB)
│ │ │ ├── dr_model_metadata.json # Feature ordering & benchmark results
│ │ │ └── held_out_test_samples.json
│ │ │
│ │ ├── outputs/features/ # Feature visualizations & evidence summaries
│ │ │ └── idrid_feature_summary.png
│ │ │
│ │ └── tests/ # Automated Verification Suites
│ │ ├── test_preprocessing.m
│ │ ├── test_quality_assessment.m
│ │ ├── test_feature_extraction.m
│ │ ├── test_model_inference.m
│ │ ├── run_all_tests.m
│ │ ├── validate_preprocessing.py
│ │ ├── validate_quality_assessment.py
│ │ ├── validate_feature_extraction.py
│ │ └── validate_model_inference.py
│ │
│ └── main.py
├── Frontend/
│ └── index.html
└── README.md
% 1. Initialize environment
run('Backend/ai-engine/matlab/setup_paths.m');
% 2. Run end-to-end inference on any new, unannotated fundus photograph
result = predict_dr('sample_fundus.png');
% Inspect results
fprintf('Status: %s\n', result.status); % 'GRADABLE' or 'UNGRADABLE'
fprintf('Predicted DR Grade: %d\n', result.predicted_grade); % 0 to 4
fprintf('Screening Confidence: %.1f%%\n', result.confidence * 100);
fprintf('Referral Recommended: %s\n', mat2str(result.referral_recommended));
fprintf('Probabilities: [%s]\n', num2str(result.class_probabilities, '%.3f '));python Backend/ai-engine/tests/validate_model_inference.pyOutput:
======================================================================
RETINO-AI MODEL INFERENCE TEST SUITE (PHASE 4 PROTOTYPE)
======================================================================
[LOAD] Loading trained model artifact -> models/saved/dr_random_forest.mat
[MODEL] Model type: random_forest | Classes: [0, 1, 2, 3, 4]
[MODEL] Features (12): ['mean_red', 'mean_green', 'mean_blue', ...]
[MODEL] Training samples: 3260 | Test samples: 815
[MODEL] Test accuracy: 66.38% | Referral accuracy: 82.45%
--- Testing predict_dr on 10 Held-Out Unseen Images (Grades 0-4) ---
[PASS] Sample 1 (APTOS e4b0df29b96f.png): Actual Grade=0 | Pred Grade=0 | Conf=68.0% | Referral=NO
[PASS] Sample 2 (APTOS b5a3ca5c0a80.png): Actual Grade=0 | Pred Grade=0 | Conf=85.6% | Referral=NO
[PASS] Sample 3 (APTOS 0684311afdfc.png): Actual Grade=1 | Pred Grade=1 | Conf=47.7% | Referral=NO
...
--- Testing Quality Gate on Intentionally Poor Quality Image ---
[PASS] Quality gate properly tripped on dark image: status=UNGRADABLE
TEST RESULT: 11 / 11 Tests Passed
>>> ALL MODEL INFERENCE AND QUALITY GATE VERIFICATIONS PASSED!
- Project: SIH26038 — Smart India Hackathon
- Datasets: APTOS 2019 Blindness Detection, Indian Diabetic Retinopathy Image Dataset (IDRiD), Digital Retinal Images for Vessel Extraction (DRIVE).