DiffuSeg: Domain-Driven Diffusion for Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Machine Learning
• Computer Vision

Background:

• Supervised machine learning for segmentation is increasing, but manual annotation is costly and error-prone.
• Deep learning models face challenges with distribution shifts between training and testing data, especially in medical imaging.
• Existing methods struggle with segmentation accuracy when target domain annotations are unavailable.

Purpose of the Study:

• To introduce DiffuSeg, a conditional diffusion model for synthesizing medical images to improve segmentation tasks.
• To address the challenge of data distribution shift in medical image segmentation.
• To enable training segmentation models without requiring new human annotations for the target domain.

Main Methods:

• Developed DiffuSeg, a novel conditional diffusion model for medical image data.
• Employed a feature factorization variational autoencoder to provide conditional information for the diffusion model.
• Utilized existing label maps and unlabeled target domain images for image synthesis.

Main Results:

• DiffuSeg demonstrated significant improvements in both image generation and segmentation accuracy compared to baselines.
• The method showed particular effectiveness in scenarios lacking target dataset annotations during training.
• Successful application to MNIST, retinal fundus vessel segmentation, and MRI heart segmentation.

Conclusions:

• DiffuSeg effectively synthesizes new images in the target domain, leveraging existing labels and unlabeled data.
• The proposed method overcomes limitations of manual annotation and distribution shifts in medical image segmentation.
• DiffuSeg offers a promising direction for training accurate segmentation models in data-scarce medical imaging scenarios.