OASIS: Generating Synthetic Skin Artifacts

Elena Sizikova, Niloufar Saharkhiz, Jana G Delfino, Aldo Badano; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2026, pp. 2993-3002

Artificial intelligence (AI) techniques have gained significant interest in recent years for automating many dermoscopic image analysis tasks. The presence of artifacts in patient datasets, either inherent to skin images (e.g., hair or blood vessels) or those introduced during image acquisition (e.g., calibration charts, rulers, dark frames) has challenged the application of these techniques since artifacts can lead to decreased performance. Furthermore, data-driven methods used to synthesize data to address patient data limitations are themselves affected by these artifacts. We present OASIS, a knowledge-based, tunable simulator for generating synthetic skin images with artifacts in the object domain prior to imaging. We describe how to simulate commonly occurring clinical artifacts, including calibration charts, rulers, dark frames, hair, and blood vessels. We then demonstrate that OASIS can be used to generate controllable, artifact-present images with pixel-level mask annotations and artifact-free counterparts, and show how the resulting images can be used for augmenting limited patient datasets in downstream tasks.