DDI2: Diverse Dermatology Images- 2

The DDI-2 dataset consists of 665 clinical photos corresponding to 169 unique biopsy-proven diagnoses from 550 patients who self-identified as Asian and were evaluated at Stanford Dermatology from 2012-2022. Patient IDs allow the user to identify which images came from the same patient. Diagnoses, anatomical location of the photo, Fitzpatrick skin type, and the presence of artifacts associated with routine clinical workflows (e.g., marker and ruler) were confirmed by a senior dermatology resident with four years of dermatology training and a board-certified attending with seven years of experience following board certification.

DDI-2 is the first real-world dataset of photos from self-identified Asian American patients with expert-curated, histopathologically-confirmed diagnoses and Fitzpatrick skin type. Existing datasets have their own limitations (ISIC contains mostly images of lighter skin; Fitzpatrick-17k is diverse but has significant label noise; Diverse Dermatology Images [DDI], which was previously curated by our group, does not contain images from self-identified Asian patients; other datasets of the Asian population are mostly private and from overseas, mostly South Korea). DDI-2 can be used to augment and evaluate model performance in the Asian American population.