A Plug-and-Play Diffusion-Styled Conversion Model for Domain Discrepancies in Medical Image Segmentation

Abstract

Accurate segmentation is a crucial step in medical image analysis. However, models trained on one dataset often suffer from performance degradation when directly applied to a different domain with a different data distribution due to domain discrepancies. To address this issue, we introduce a novel plug-and-play diffusion-styled domain conversion method, named Diffusion-Styled Domain Conversion (DSDC). Our DSDC framework formulates domain conversion as a gradual transition process from domain A to domain B, inspired by the diffusion process. Unlike traditional diffusion models that iteratively remove noise, our model progressively shifts the data distribution from domain A to domain B. We tested our DSDC method on two retinal vessel segmentation datasets, combining it with seven different segmentation models, and observed an average improvement of 20% in the DICE score and a reduction of 10 mm in the average ASD.