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

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Computer Vision

Background:

• Generative dataset expansion methods are crucial for addressing data scarcity in dermoscopic image segmentation.
• Existing two-stage synthesis strategies often incur high computational costs due to complex designs and additional learnable components.

Purpose of the Study:

• To propose SCCS-Diff, a novel, efficient, and simple one-stage generative dataset expansion method for dermoscopic images.
• To leverage self-conditioning strategies within a Latent Diffusion Model (LDM) paradigm for improved data synthesis.

Main Methods:

• SCCS-Diff employs a one-stage synthesis framework integrating a self-conditioning strategy.
• It utilizes a variational autoencoder for trajectory correction in the reversed diffusion process.
• The method synthesizes aligned dermoscopic image-mask pairs efficiently without complex conditioning designs.

Main Results:

• SCCS-Diff demonstrates effectiveness in synthesizing high-fidelity dermoscopic image-mask pairs.
• Comparisons and ablation studies on ISIC-2016, 2017, and 2018 datasets validate the method's performance.
• The approach avoids additional training costs and complex design elements inherent in other methods.

Conclusions:

• SCCS-Diff offers an effective solution for alleviating data scarcity in medical imaging datasets, specifically dermoscopy.
• The proposed method provides a computationally efficient and simpler alternative for generative dataset expansion.
• This work highlights the potential of self-conditioning diffusion models for medical image segmentation tasks.