A skin lesion hair mask dataset with fine-grained annotations

The largest publicly available skin lesion hair segmentation mask dataset created by carefully annotating 500 copyright-free CC0 licensed dermoscopic images collected from ISIC 2018 dataset [1]. The dataset is organized into three folders namely dermoscopic_image, hair_mask, and overlay. The dermoscopic_image folder contains 500 handpicked dermoscopic images covering different hair patterns. We retained the original names of the image files from the primary image source. The hair_mask folder contains a binary segmentation mask for each of the images of the dermoscopic_image folder. In a segmentation mask image, white pixels represent skin hair and black pixels represent background. The overlay folder contains hair mask images superimposed on the original dermoscopic images. We provided the superimposed images for easy public verification so that, other people can report any annotation mistakes and contribute to improving the dataset. Images in the hair_mask and overlay folders share the same names as the primary images in the dermoscopic_image folder. additional_materials folder contains codes and additional materials used for preparing the dataset.

additional_materials folder contents:
- Inside the unet folder the U-net [2] model is defined in model.pyfile, unet training is performed using the unet_training.ipynb python notebook file. The task of predicting initial masks for the dermoscopic images is done using the predict_mask.ipynbfile.
- The codes used for binarizing mask, making it transparent and creating image collage are available in the check_annotation.ipynbfile.
- Video demonstration of the hair mask editing process is available in the mask_editing_process.mp4 file.

References
[1] Codella N, Rotemberg V, Tschandl P, Celebi ME, Dusza S, Gutman D, et al. Skin Lesion Analysis Toward Melanoma Detection 2018: A Challenge Hosted by the International Skin Imaging Collaboration (ISIC) 2019. https://doi.org/10.48550/arxiv.1902.03368.
[2] Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation. In: Navab N, Hornegger J, Wells WM, Frangi AF, editors. Med. Image Comput. Comput. Interv. -- MICCAI 2015, Cham: Springer International Publishing; 2015, p. 234–41. https://doi.org/10.1007/978-3-319-24574-4_28

本数据集为当前公开可用的规模最大的皮肤病变毛发分割掩码数据集，构建方式为：从ISIC 2018数据集[1]中遴选500张遵循CC0授权的无版权皮肤镜图像，并经人工精心标注完成。
数据集分为三个子文件夹，分别为dermoscopic_image（皮肤镜图像）、hair_mask（毛发掩码）与overlay（叠加图）。其中dermoscopic_image（皮肤镜图像）文件夹包含500张精心甄选、涵盖多种毛发形态的皮肤镜图像，且保留了原始图像来源的文件名。hair_mask（毛发掩码）文件夹包含与dermoscopic_image（皮肤镜图像）文件夹中每张图像对应的二值分割掩码：在分割掩码图像中，白色像素代表皮肤毛发，黑色像素代表背景。overlay（叠加图）文件夹存储将毛发掩码叠加至原始皮肤镜图像后生成的可视化结果，我们提供此类叠加图以方便公众核验，便于他人反馈标注误差并助力本数据集的迭代优化。hair_mask（毛发掩码）与overlay（叠加图）文件夹内的图像与dermoscopic_image（皮肤镜图像）文件夹内的原始图像同名。
additional_materials（附加材料）文件夹包含用于数据集构建的代码与配套素材：
- 在unet子文件夹中，U-net（U-Net）模型的定义代码存放于model.py文件，U-net模型训练通过unet_training.ipynb Python交互式笔记本文件完成；针对皮肤镜图像的初始掩码预测任务由predict_mask.ipynb文件实现。
- 用于实现掩码二值化、设置透明度及生成图像拼贴的代码收录于check_annotation.ipynb文件中。
- mask_editing_process.mp4文件为毛发掩码编辑流程的视频演示素材。

参考文献
[1] Codella N, Rotemberg V, Tschandl P, Celebi ME, Dusza S, Gutman D, 等. 面向黑色素瘤检测的2018年皮肤病变分析：国际皮肤影像协作组（ISIC）主办挑战赛[EB/OL]. 2019. https://doi.org/10.48550/arxiv.1902.03368.
[2] Ronneberger O, Fischer P, Brox T. U-Net：用于生物医学图像分割的卷积网络[C]//Navab N, Hornegger J, Wells WM, Frangi AF, 编者. 医学图像计算与计算机辅助干预——MICCAI 2015. 沙姆：施普林格国际出版; 2015: 234–241. https://doi.org/10.1007/978-3-319-24574-4_28