DRP Segmentation — Diabetic Retinopathy Screening Prototype

Open NB01 In Colab

An AI-assisted diabetic retinopathy (DR) screening prototype that detects and grades DR from fundus images, highlights suspicious retinal regions, and provides interpretable outputs for clinical decision support.

DRP_segmentation/
├── notebooks/
│ ├── 01_disease_segmentation.ipynb # DR grading preprocessing — APTOS 2019
│ ├── 02_vessel_segmentation.ipynb # Vessel segmentation preprocessing — DRIVE
│ └── 03_binary_classification_filter.ipynb # Binary DR filter — DR Resized
├── src/
│ └── engine/
│ └── image_preprocessing.py # Core preprocessing pipeline
├── data/
│ ├── raw/ # Downloaded datasets (gitignored)
│ └── processed/ # Preprocessed .npz files (gitignored)
└── README.md

All notebooks share the same core pipeline (src/engine/image_preprocessing.py):

1. Load — Read fundus image (PNG/JPEG/TIFF)
2. Resize — Standardize to ×ばつ512
3. CLAHE — Contrast Limited Adaptive Histogram Equalization for vessel/lesion enhancement
4. Normalize — Zero-one scaling [0, 1]
5. Save — Export as compressed .npz with image array + metadata

Before running any notebook, add your Kaggle credentials to Colab Secrets (🔑 icon):

Get your token from: kaggle.com/settings → API → Create New Token

1. Open notebook in Colab (use badge above)
2. Runtime → Run all (Ctrl+F9)
3. Run Cell 11 separately to save .npz files to Google Drive

• DR Grading: Quadratic Weighted Kappa (QWK), AUC-ROC, Accuracy per grade
• Segmentation: Dice coefficient, IoU, Sensitivity, Specificity
• Classification: AUC-ROC, F1-score, Sensitivity, Specificity
• Explainability: Grad-CAM heatmaps for lesion localization

• Notebook 06 — Grad-CAM Explainability
• Notebook 07 — Clinical Interface (Gradio/Streamlit)