retino-cnn

An app using CNN to determine severity od diabetic retinopathy.

Environment

Dev environment is managed using poetry: https://python-poetry.org.

poetry install --with dev

Dataset

About dataset

Resized 2015 & 2019 Diabetic Retinopathy Detection from Kaggle: https://www.kaggle.com/datasets/c7934597/resized-2015-2019-diabetic-retinopathy-detection
Dataset contains 92 404 images of human retina including:

• 67 156 healthy
• 6 583 mild cases
• 14 160 moderate cases
• 2 288 severe cases
• 2 217 proliferative cases

In total about 72.5% of dataset are healthy retinas and 27.5% are not-healthy with varying severity.
I simplify this to binary classification problem: healthy vs not-healthy.

Download dataset

To download the dataset, first create a personal Kaggle API token under
Settings -> Account -> API https://www.kaggle.com/settings
and store it in ${KAGGLE_CONFIG_DIR}/kaggle.json directory (or ~/.kaggle/kaggle.json which is default for Kaggle API).

Next run make command:

make all

to input environment variables, download the dataset and split it into train, validation and test sets.
Alternatively, you can run each step separately:

make input-env # Create input environment file
make download-dataset # Download dataset from Kaggle (with optional unzipping)
make rename-files # Rename files to remove whitespaces and parentheses
make split-data # Split data into train, validation and test sets

or if you have .env file already created:

make download-rename-data # Download, rename, split