skinisic provides a trained convolutional neural network to detect the presence of four dermoscopic criteria (pigment network, negative network, milia-like cysts, streaks) within a dermoscopy image.
Our entry had the highest AUROC score on the ISBI ISIC 2017 Challenge Part 2 - Lesion Dermoscopic Feature Extraction, and ranked first place on this task.
The CNN model provided achieves a higher Jaccard Index than the CNN entry used in the challenge. We make the case that the Jaccard Index provides a more clinically meaningful measure of performance than the AUROC score. More details can be found in,
Jeremy Kawahara and Ghassan Hamarneh, “Fully convolutional neural networks to detect clinical dermoscopic features,” IEEE Journal of Biomedical and Health Informatics, vol. 23, no. 2, pp. 578–585, 2019. [DOI] [PDF]
skinisic is a Python module that relies on Keras.
You can see the dependencies and versions tested on here.
To use skinisic,
- Navigate to your desired directory (e.g.,
/projectsfor this example) and open terminal. - Clone this repository to your local machine:
git clone https://github.com/jeremykawahara/skinisic.git - Download the trained model and save to disk (e.g,.
/projects/skinisic/notebooks/data):
https://github.com/jeremykawahara/skinisic/releases/download/v0.0.1/isic2017-part2_vgg_f1-batch_aug.h5 - Navigate to the
skinisicdirectory and run the minimal example (may take a few seconds):
cd skinisic
python minimal_example.py 'notebooks/data/isic2017-part2_vgg_f1-batch_aug.h5'
You should see the following output:
You can find a more comprehensive example here.
If you find this code or model helpful, please consider citing our work:
@article{Kawahara2019,
author = {Kawahara, Jeremy and Hamarneh, Ghassan},
doi = {10.1109/JBHI.2018.2831680},
issn = {21682194},
journal = {IEEE Journal of Biomedical and Health Informatics},
number = {2},
pages = {578--585},
title = {Fully convolutional neural networks to detect clinical dermoscopic features},
volume = {23},
year = {2019}
}