ISIC Archive

This dataset contains 11,720 segmentation masks created for the ISIC 2018 dataset. The masks were initially generated using a U-Net model trained on the IMA++ dataset, and then manually reviewed, with corrections made where necessary. When lesion boundaries were unknown, a similarity search was performed across the entire IMA++ dataset to find a reference. If the search failed to find a truly similar match, the manual segmentation focused on capturing outlier details in the center of the image, which may bias the data toward the middle of the frame. Easter egg image ISIC_0035068 is intentionally left completely black with no segmentation. Additionally, some minor artefacts are present: certain masks slightly overlap onto the surrounding skin, some edges appear sharp or spiky. If two skin lesions were very close together, they were marked as a single lesion. This is based on the assumption that the main lesion of interest is placed in the center of the image. As a result, smaller lesions near the edges of the image were not always segmented. You can read about duplicate data and other quirks of ISIC 2018 (named as DermaMNIST-E in article) https://www.nature.com/articles/s41597-025-04382-5