Diabetic Retinopathy Cascade Classification

---

📌 Overview

This project implements a multi-stage cascaded deep learning framework for Diabetic Retinopathy (DR) classification using a pre-trained ResNet50 backbone.
The approach improves both early-stage and advanced-stage DR detection by:

• Grouping visually similar DR stages.
• Refining classification in multiple cascade stages.

The model is trained and evaluated using two public datasets:

• APTOS 2019 Blindness Detection
• Diabetic Retinopathy Resized Dataset

By combining datasets and applying a Smooth Data Augmentation strategy, this system effectively addresses class imbalance and enhances detection of subtle early-stage DR cases.

---

✨ Key Features

• Cascaded ResNet50 architecture for staged DR classification.
• Dual dataset integration for increased diversity and robustness.
• Structured augmentation pipeline (flip, brightness, saturation) for class balancing.
• Detailed evaluation metrics: class-wise F1-scores, precision, and recall.
• Improved early-stage sensitivity, critical for real-world screening.

---

📂 Repository Structure

├── Augmentation/                # Data augmentation scripts
├── Data Management/              # Dataset organization & splitting
├── Data Preprocessing/           # Image preprocessing scripts
├── Data/                         
│   ├── Raw/                      # Raw dataset info 
│   │   └── README.md              # Dataset sources & original distribution
│   ├── Processed/                 # Processed dataset info
│       └── README.md              # Post-preprocessing & augmentation counts
├── Docs/                          # Full project report & documentation
├── Model/                         # Model training notebook
├── Results/                       # Performance metrics & visualizations
├── Weights/                       # Trained model weights 
├── LICENSE
└── README.md