Exploring Self-Conditioning Co-Sample Strategy of Diffusion Models in Dermoscopic Images

Generative dataset expansion methods can effectively alleviate the scarcity of data in dermoscopic image segmentation but commonly employ a two-stage synthesis strategy that contains additional learnable components and complex design, which results in high computational resource costs. Diffusion models utilizing a self-conditioning strategy have shown strong potential for efficiently reusing priors in the pipeline without relying on excessively complicated conditioning designs. Inspired by this, we propose a dataset expansion method called SCCS-Diff. It utilizes a simple and efficient one-stage synthesis framework and introduces a self-conditioning strategy based on the Latent Diffusion Model paradigm. Our proposed SCCS-Diff can synthesize the highly aligned dermoscopic image-mask pairs at once by efficiently multiplexing the variational autoencoder to accomplish the trajectory correction of the reversed process, thus effectively avoiding additional training costs and complex design. The comparisons with previous methods and the ablations on the ISIC-2016, 2017 and 2018 datasets demonstrate the effectiveness of SCCS-Diff in fidelity and data pair matching. We expect that SCCS-Diff will provide an effective solution to alleviate the scarcity of medical imaging datasets.