Abstract
Machine learning classification algorithms have emerged as promising tools to support the early detection of skin cancers. Existing algorithms typically assess malignancy of skin lesions based on a single skin image. This is in contrast with how clinicians integrate information from their physical examination, comparing multiple skin lesions of an individual and changes in lesions over time. Including contextual information could greatly enhance machine learning algorithms. However, contextual information in skin image datasets is predominantly scarce and inconsistent. Additionally, a dataset containing images of the same lesion across multiple time points and varying resolutions is also lacking. To address these gaps, we present a comprehensive dataset derived from skin monitoring of 480 study participants recruited from a general population sample (n = 196) and a high-risk for melanoma cohort (n = 284). This dataset includes images of 250,162 skin lesions obtained from three-dimensional total body imaging (tile images), along with corresponding dermoscopic images of 9,389 lesions. For 340 of the participants, longitudinal tile and dermoscopic images (ranging from 2 to 7) are provided.