The Effects of Skin Lesion Segmentation on the Performance of Dermatoscopic Image Classification

This repository contains the derived classification probability vectors (from both sets of experiments) and also automatically created segmentation masks (SkinUNet, SkinFPN+, and SkinLinkNet) that were used in our study. The classification probability vectors and the segmentation masks of the ISIC 2017 test set images can be found in the respective folders.

Files description:

• predictions folder: contains classification predictions for both sets of experiments. The classification prediction vectors for the first set of experiments are inside the "using manual masks" folder, the classification prediction vectors for the second set of experiments are inside the "using automatic masks (SkinLinkNet)" folder and the classification prediction vectors for the reference dataset is inside the "using no masks (reference dataset)" folder. Additionally, the classification probability vectors for the SkinFPN+ and SkinUNet for the dilated cropped scenario can be found in "using automatic masks (SkinFPN+)" folder and "using automatic masks (SkinUNet)" folder, respectively.

• Segmentation_masks folder: contains the automatically created segmentation masks (from SkinUNet, SkinFPN+, and SkinLinkNet) for all 600 test images of the ISIC 2017 test set.

• image_ids.pdf: Contains two tables. The first table shows the image IDs (from the ISIC 2017 test set) of the images that were only classified correctly when the reference dataset was used but were incorrectly classified when the dilated cropped dataset was used. The second table shows the image IDs (from the ISIC 2017 test set) of the images that were only classified correctly when the dilated cropped dataset was used but were incorrectly classified when the reference dataset was used.

Useful links to perform statistical tests:

• https://www.medcalc.org/manual/comparison_of_roc_curves.php (for AUC comparison)
• https://ch.mathworks.com/help/stats/testcholdout.html (for McNemar statistical test)
• https://ch.mathworks.com/help/stats/signrank.html (for Wilcoxon signed-rank test method)

Citations

The paper describing our methodology is accepted in the Computer Methods and Programs in Biomedicine journal and can be found here:

BibTex entry:

@article{MAHBOD2020105725,
title = "The Effects of Skin Lesion Segmentation on the Performance of Dermatoscopic Image Classification",
journal = "Computer Methods and Programs in Biomedicine",
pages = "105725",
year = "2020",
issn = "0169-2607",
doi = "https://doi.org/10.1016/j.cmpb.2020.105725",
author = "Amirreza Mahbod and Philipp Tschandl and Georg Langs and Rupert Ecker and Isabella Ellinger",
keywords = "Skin cancer, dermatoscopy, medical image analysis, deep learning, effect of segmentation on classification",
}

The description of the utilized classification workflow in our paper can be found in the following article:

BibTex entry:

@article{MAHBOD2020105475,
title = "Transfer Learning Using a Multi-Scale and Multi-Network Ensemble for Skin Lesion Classification",
journal = "Computer Methods and Programs in Biomedicine",
pages = "105475",
year = "2020",
issn = "0169-2607",
doi = "https://doi.org/10.1016/j.cmpb.2020.105475",
author = "Amirreza Mahbod and Gerald Schaefer and Chunliang Wang and Georg Dorffner and Rupert Ecker and Isabella Ellinger",
keywords = "Skin cancer, dermoscopy, medical image analysis, deep learning, image resolution, image cropping, transfer learning",
}

License

This project is licensed under the Apache (version 2.0) License.