DR-ENb0

Diabetic retinopathy severity grading on APTOS 2019 fundus images using EfficientNet-B0, SMOTE-based class balancing, focal-loss variants, and Grad-CAM-based qualitative explainability.

Requirements

• Python 3.13
• uv

Setup

git clone https://github.com/yassyass2/DR-ENb0.git
cd DR-ENb0
uv sync

Project Layout

capstone-project/
├── configs/                 # Training configuration
├── models/                  # Saved .keras models and model notebooks
│   └── preprocessed/        # Saved .npz dataset archive for inference
├── scripts/                 # Standalone utility scripts
├── src/dr_grading/          # Package code
└── tests/                   # Lightweight structural tests
pyproject.toml
uv.lock

Main Commands

Preprocess the dataset:

uv run dr-grading preprocess

Train with the default config:

uv run dr-grading train

List saved models:

uv run python capstone-project/scripts/predict_saved_model.py --list-models

Run one saved model on one preprocessed sample:

uv run python capstone-project/scripts/predict_saved_model.py --model 1 --split test --index 0

Interactive mode:

uv run python capstone-project/scripts/predict_saved_model.py

Show script help with the current discovered model names:

uv run python capstone-project/scripts/predict_saved_model.py --help

Download the preprocessed .npz archive from Google Drive:

uv run python capstone-project/scripts/download_preprocessed_npz.py

Saved Models

The capstone-project/models/ folder currently contains these final saved .keras models:

1. SMOTE.keras
2. SMOTE_+_Focal_Loss_IoU.keras
3. SMOTE_+_Focal_loss.keras
4. SMOTE_IoU.keras

The prediction script accepts:

• a model number from the discovered list
• a model filename such as SMOTE_IoU.keras
• a full filesystem path to a .keras file

Preprocessed Inference Data

The saved-model prediction workflow uses:

capstone-project/models/preprocessed/dr_preprocessed_data.npz

It loads these arrays:

• X_train, y_train
• X_val, y_val
• X_test, y_test

By default the prediction script uses the .npz archive above, but it can also load a directory of split .npy arrays.

If the archive is missing locally, predict_saved_model.py now downloads it automatically from the configured Google Drive link before running inference.

You can also download it manually with:

uv run python capstone-project/scripts/download_preprocessed_npz.py

Prediction Output

For a selected image, the script prints:

• true label and class name
• predicted label and class name
• confidence of the top prediction
• whether the prediction was correct
• prediction margin between the top two classes
• entropy of the probability distribution
• per-class probabilities

Dataset

APTOS 2019 Blindness Detection

Labels are the diabetic retinopathy severity grades 0 through 4.

Contributors

• Adel
• Mark
• Musab
• Ozeir
• Yassine