Diabetic Retinopathy Grading from Fundus Images using CNN

Diabetic retinopathy is a leading cause of preventable blindness: chronic high blood sugar damages the small vessels of the retina, progressing silently from microaneurysms to hemorrhages, exudates and finally fragile new blood vessels that can bleed into the eye. Systematic fundus-photograph screening catches it early, but there are far fewer trained ophthalmologists than diabetic patients who need screening. This project builds that screening idea as an honest educational prototype: an EfficientNetB0 network fine-tuned on the public APTOS 2019 Blindness Detection set grades a fundus photograph into the five International Clinical Diabetic Retinopathy stages, evaluated with quadratic weighted kappa and per-class AUC rather than a single accuracy number. The interactive demo lets you select a sample fundus image, run the grading, and inspect the predicted stage with per-class probabilities and a lesion-highlight overlay — while every page of the project states plainly that this is a teaching build, not a diagnostic device.

1. The 3,662 labeled APTOS 2019 fundus images are circular-cropped to remove black borders, CLAHE-enhanced, resized to 224×224×3 and normalized with ImageNet statistics.
2. Stratified 80/10/10 train/validation/test splits preserve the stage ratio; augmentation (rotation, flips, brightness jitter) is applied during training.
3. EfficientNetB0 (ImageNet weights) is fine-tuned with a custom head — GAP → Dense(256, ReLU) → Dropout(0.4) → Dense(5, softmax) — using weighted categorical cross-entropy and Adam (lr 1e-4).
4. Training runs with early stopping on validation quadratic weighted kappa; the best checkpoint is kept and the curves recorded.
5. The held-out test split is evaluated once: quadratic weighted kappa, per-class AUC and the confusion matrix are generated for the report.
6. In the web demo, a selected fundus image is preprocessed identically and passed through the saved weights; the app shows the predicted stage, probability bars and the lesion overlay. All outputs are labeled illustrative — not a diagnosis.

Tech stack:

• Python 3, TensorFlow/Keras
• EfficientNetB0 (ImageNet pretraining, fine-tuned)
• OpenCV, NumPy, scikit-learn (preprocessing & metrics)
• Jupyter Notebook (training & evaluation)
• HTML5 + JavaScript (interactive grading demo)
• APTOS 2019 Blindness Detection set (Kaggle, 3,662 images)

Dataset & model details

• Dataset: APTOS 2019 Blindness Detection (Kaggle, Asia Pacific Tele-Ophthalmology Society) — 3,662 labeled retinal fundus photographs across 5 ICDR stages: No DR 1,805 / Mild 370 / Moderate 999 / Severe 193 / Proliferative 295; plus 1,928 test images with hidden labels.
• Task: 5-class image classification; input = 224×224×3 RGB fundus image, output = probability distribution over the 5 DR stages.
• Model: EfficientNetB0 backbone (ImageNet-pretrained) fine-tuned end-to-end; head = GlobalAveragePooling → Dense(256, ReLU) → Dropout(0.4) → Dense(5, softmax). Approximately 4.1M backbone parameters plus ~330k in the head. Weighted categorical cross-entropy, Adam (lr 1e-4), early stopping on validation quadratic weighted kappa.
• Metrics: Quadratic weighted kappa ≈ 0.80, mean per-class AUC ≈ 0.88, referable-DR sensitivity ≈ 0.90 (design targets for the built-to-order training run), confusion matrix. No figure is claimed as measured until the training run is executed for the order; the report records the achieved values.

• [EfficientNetB0 transfer learning] An ImageNet-pretrained EfficientNetB0 backbone fine-tuned on APTOS 2019 with a custom classification head — global average pooling, Dense(256, ReLU), Dropout(0.4) and a 5-way softmax — with every layer choice documented.
• [Interactive fundus grading demo] Select a sample fundus photograph and run the CNN grading; the app shows the predicted stage, per-class probability bars, a lesion summary and a clinical note in a clean medical-dashboard interface.
• [Lesion-highlight overlay] A toggleable attention overlay marks hemorrhages, exudates and neovascular regions the model weighted most — the standard explainability visual for retinal imaging, labeled illustrative in the demo.
• [Real APTOS 2019 dataset] 3,662 labeled fundus photographs across the five ICDR stages (No DR 1,805 / Mild 370 / Moderate 999 / Severe 193 / Proliferative 295) — a real, citable public benchmark.
• [Competition-grade evaluation] Quadratic weighted kappa plus per-class AUC instead of a lone accuracy number, with the severe class imbalance handled by weighted loss — the way retinal grading is actually evaluated.
• [Full training notebook] Circular crop, CLAHE contrast enhancement, augmentation, model definition, training with early stopping on validation kappa, and evaluation in one reproducible notebook.
• [Training curves and attention maps] Validation kappa and loss curves plus Grad-CAM-style attention examples across the five stages, included in the report.
• [Medical-limits documentation] The report states clearly that the build is an educational prototype, not a certified diagnostic device, with the validation steps a real deployment would need.

• Complete training & evaluation Jupyter notebook (APTOS loading, CLAHE preprocessing, augmentation, fine-tuning, kappa evaluation)
• Trained EfficientNetB0 model file with preprocessing code
• Interactive fundus-grading web demo (samples, probability bars, lesion overlay, clinical notes)
• Evaluation plots: training curves, per-class metrics table, sample attention maps
• Project report PDF (background, dataset, architecture rationale, methodology, honest results section, medical-AI limits)
• PPT presentation for final review
• Viva Q&A preparation document (transfer learning, quadratic weighted kappa, class imbalance, Grad-CAM, why this is not a diagnostic device)