MedSegDiff

Why we included this project

MedSegDiff treats medical image segmentation as a denoising problem: a diffusion probabilistic model generates organ and lesion boundaries from scans rather than classifying each pixel with a conventional encoder-decoder. The repository includes data loaders and working scripts for training, sampling, and evaluation on two real benchmarks, ISIC skin-lesion segmentation and BraTS brain-tumor segmentation, plus guidance on extending it to your own data with a custom loader. The V2 branch swaps in a transformer backbone and reports better accuracy and stability, while the bundled DPM-Solver cuts sampling from 1000 diffusion steps to around 20, making experiments much cheaper. It's a research codebase from a paper named an AAAI most influential paper, so expect to adapt it rather than deploy it as-is. For teams comparing diffusion-based segmentation with classical models, this is a solid, actively maintained reference point.