Skin lesion classification

Skin lesion binary classification using Keras and the ISIC 2020 dataset.

Setup the Python environment and download the dataset

All the required packages can be installed with pip:

pip install -r requirements.txt

It's better to use a virtual env to prevent version conflicts between packages.

Then you'll have to download the ISIC 2020 train dataset as well as the metadata as CSV files. This can be done automatically with the setup_dataset.sh script:

./setup_dataset.sh

How to train a model

./train.py [--remove-artifacts] [--segmentation] [--checkpoint-folder FOLDER] [--epochs EPOCHS] [--batch-size BATCH_SIZE]

Available options:

• --remove-artifacts to perform artifacts removal with morphological closing
• --segmentation to segment the images with the K-means algorithm
• --checkpoint-folder folder to save model checkpoints and final model. Default is checkpoints/
• --epochs maximum number of epochs to train for. Default is 300
• --batch-size batch size to use for training. Default is 256
• --notifier-prefix header to send in Telegram messages when sending training progress
• --help, -h show available options

How make a prediction using a trained model

Pretrained models can be downloaded in the Releases page.

./test.py image model [--segment] [--remove-artifacts]

image must be the path to the image to evaluate
model must be the path of the saved Keras model

The output is the probability that the input image is malignant.

---

Project report