SkinDiff: A Novel Data Synthesis Method Based on Latent Diffusion Model for Skin Lesion Segmentation

Recent advancements in skin lesion segmentation focus on model architecture refinement, but they are hindered by limited datasets and resource-consuming manual annotation. To address this issue, this paper proposes SkinDiff, a novel framework for training data expansion. Derived from the Latent Diffusion Model, we utilize two steps, the Generating Foreground and the Outpainting Background techniques, to synthesize high-fidelity labeled image samples to enhance model training for skin lesion segmentation. The Generating Foreground technique produces realistic skin lesions and obtains their corresponding masks, and then the Outpainting Background technique fills in the normal skin around the lesions to obtain full images. Experimental results show that expanding the skin lesion dataset using the proposed method can significantly improve the performance of the segmentation model.