skin-lesion-classification

Repository for skin lesion classification task

Setup:

Clone the repository

git clone https://github.com/pavelg0t/skin-lesion-classification.git
cd skin-lesion-classification/

Setup virtual environment with venv

python3 -m venv venv
source venv/bin/activate

Install necessary packages (Python version used: Python 3.8.10)

pip install -r requirements.txt

Setup pre-commit hooks

pre-commit install

Initially run the pre-commit hooks on all of the files in the repo:

pre-commit run --all-files

Setup WandB for experiment tracking

Install the pip package (if not already installed with the requirements.txt file)

pip install wandb --upgrade
wandb login

Go to https://wandb.ai/authorize, login, copy the key and paste it in the terminal. The API key will be automatically saved to /home/<user>/.netrc.

Download the data (via Kaggle)

Set up Kaggle API CLI Install the pip package (if not already installed with the requirements.txt file)

pip install kaggle

Create an API Token

Go to (https://www.kaggle.com/<username>/account) and select 'Create New API Token'. Then move this JSON file to ~/.kaggle/kaggle.json:

mv <path_to_file>/kaggle.json ~/.kaggle/kaggle.json

To ensure that other computer users do not have read access to your credentials:

chmod 600 ~/.kaggle/kaggle.json

Or export the credentials as env vars:

export KAGGLE_USERNAME="username"
export KAGGLE_KEY="key"

Download the datasets

HAM10000

kaggle datasets download -p data/HAM10000 --unzip kmader/skin-cancer-mnist-ham10000

ISIC2020

bash setup_isic2020.sh

Train

Binary classification with timm model (xcit nano)

python3 train.py First_Project --m_name xcit_nano_12_p16_224 --dataset ISIC2020 --multiclass false --transform timm_original --num_workers 4 --batch_size 8 --n_epoch 2

References

• Rotemberg, V., Kurtansky, N., Betz-Stablein, B., Caffery, L., Chousakos, E., Codella, N., Combalia, M., Dusza, S., Guitera, P., Gutman, D., Halpern, A., Helba, B., Kittler, H., Kose, K., Langer, S., Lioprys, K., Malvehy, J., Musthaq, S., Nanda, J., Reiter, O., Shih, G., Stratigos, A., Tschandl, P., Weber, J. & Soyer, P. A patient-centric dataset of images and metadata for identifying melanomas using clinical context. Sci Data 8, 34 (2021). https://doi.org/10.1038/s41597-021-00815-z
• Tschandl, P., Rosendahl, C. & Kittler, H. The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. Sci. Data 5, 180161 doi:10.1038/sdata.2018.161 (2018).