This repository contains the implementation of a novel medical image segmentation method that combines diffusion models and a Structure-Preserving Network for structure-aware one-shot image stylization. The proposed approach aims to mitigate domain shifts caused by variations in imaging devices, acquisition conditions, and patient-specific attributes, which often challenge the accuracy of medical image segmentation.
You can use link OSASIS(https://github.com/hansam95/OSASIS?tab=readme-ov-file#prepare-training) to transfer the style of the image. It is recommended to use the resize_and_stave_images function.
In the style transfer of polyp segmentation, we recommend training for 50 iterations, while in the style transfer of skin lesion segmentation, we recommend training for 30 iterations.
We propose a novel approach that integrates diffusion-based style transfer with image segmentation to enhance the robustness of medical image segmentation under domain shifts.
If our ideas are helpful to you, please cite our article:
@article{Jie2024Structure,
title={Structure-Aware Stylized Image Synthesis for Robust Medical Image Segmentation},
author={Bao, Jie and Zhou, Zhixin Zhou and Li, Wen Jung and Luo, Rui},
journal={arXiv preprint arXiv:2412.04296},
year={2024}
}