Title: IMA++: A Multi-Annotator Dermoscopic Skin Lesion Segmentation Dataset from the ISIC Archive

Abstract:

While multi-annotator medical image segmentation represents a critical area of inquiry, the high cost of gathering annotated data remains a significant barrier. Dermoscopy enhances the visibility of morphological structures that are often invisible in standard clinical photography, benefiting both human specialists and artificial intelligence models. Despite this utility, there is currently a lack of large-scale, publicly accessible skin lesion segmentation (SLS) datasets that include annotator labels for dermoscopic imagery.

To address this gap, we present ISIC MultiAnnot++, a comprehensive public dataset derived from the ISIC Archive. This collection stands as the largest publicly available SLS dataset, comprising 17,684 segmentation masks across 14,967 dermoscopic images. Notably, 2,394 of these images feature between two and five distinct segmentations each. The dataset is enriched with segmentation metadata, such as the annotators’ proficiency levels and the specific tools used, thereby facilitating research into annotator-specific preference modeling and metadata analysis. Additionally, we offer an examination of the dataset’s characteristics, alongside curated data partitions and consensus segmentation masks.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC