MedSegDiff a Diffusion Probabilistic Model (DPM) based framework for Medical Image Segmentation. The algorithm is elaborated in our paper MedSegDiff: Medical Image Segmentation with Diffusion Probabilistic Model and MedSegDiff-V2: Diffusion based Medical Image Segmentation with Transformer.
Diffusion Models work by destroying training data through the successive addition of Gaussian noise, and then learning to recover the data by reversing this noising process. After training, we can use the Diffusion Model to generate data by simply passing randomly sampled noise through the learned denoising process.In this project, we extend this idea to medical image segmentation. We utilize the original image as a condition and generate multiple segmentation maps from random noises, then perform ensembling on them to obtain the final result. This approach captures the uncertainty in medical images and outperforms previous methods on several benchmarks.
 |
 |
|---|---|
| MedSegDiff-V1 | MedSegDiff-V2 |
- 22-11-30. This project is still quickly updating. Check TODO list to see what will be released next.
- 22-12-03. BraTs2020 bugs fixed. Example case added.
- 22-12-15. Fix multi-gpu distributed training.
- 22-12-16. DPM-Solver ✖️ MedSegDiff DONE 🥳 Now DPM-Solver is avaliable in MedsegDiff. Enjoy its lightning-fast sampling (1000 steps ❌ 20 steps ⭕️) by setting
--dpm_solver True. - 22-12-23. Fixed some bugs of DPM-Solver.
- 22-01-31. MedSegDiff-V2 will be avaliable soon 🥳 . Check our paper MedSegDiff-V2: Diffusion based Medical Image Segmentation with Transformer first.
- 22-02-07. Optimize workflow in BRATS sampling. Add dataloader for processing raw 3D BRATS data.
- 22-02-11. Fix bugs 3D BRATS data training bugs, issue 31.
- 22-03-04. Paper MedSegDiff: Medical Image Segmentation with Diffusion Probabilistic Model has been officially accepted by MIDL 2023 🥳
Python 3.8.0+
Pytorch 1.8.1+
- Download ISIC dataset from https://challenge.isic-archive.com/data/. Your dataset folder under "data_dir" should be like:
ISIC/
ISBI2016_ISIC_Part3B_Test_Data/...
ISBI2016_ISIC_Part3B_Training_Data/...
ISBI2016_ISIC_Part3B_Test_GroundTruth.csv
ISBI2016_ISIC_Part3B_Training_GroundTruth.csv
-
For training, run:
python scripts/segmentation_train.py --data_name ISIC --data_dir input data direction --out_dir output data direction --image_size 256 --num_channels 128 --class_cond False --num_res_blocks 2 --num_heads 1 --learn_sigma True --use_scale_shift_norm False --attention_resolutions 16 --diffusion_steps 1000 --noise_schedule linear --rescale_learned_sigmas False --rescale_timesteps False --lr 1e-4 --batch_size 8 -
For sampling, run:
python scripts/segmentation_sample.py --data_name ISIC --data_dir input data direction --out_dir output data direction --model_path saved model --image_size 256 --num_channels 128 --class_cond False --num_res_blocks 2 --num_heads 1 --learn_sigma True --use_scale_shift_norm False --attention_resolutions 16 --diffusion_steps 1000 --noise_schedule linear --rescale_learned_sigmas False --rescale_timesteps False --num_ensemble 5
In default, the samples will be saved at ./results/
- Download BRATS2020 dataset from https://www.med.upenn.edu/cbica/brats2020/data.html. Your dataset folder should be like:
data
└───training
│ └───slice0001
│ │ t1.nii.gz
│ │ t2.nii.gz
│ │ flair.nii.gz
│ │ t1ce.nii.gz
│ │ seg.nii.gz
│ └───slice0002
│ │ ...
└───testing
│ └───slice1000
│ │ t1.nii.gz
│ │ t2.nii.gz
│ │ flair.nii.gz
│ │ t1ce.nii.gz
│ └───slice1001
│ │ ...
-
For training, run:
python scripts/segmentation_train.py --data_dir (where you put data folder)/data/training --out_dir output data direction --image_size 256 --num_channels 128 --class_cond False --num_res_blocks 2 --num_heads 1 --learn_sigma True --use_scale_shift_norm False --attention_resolutions 16 --diffusion_steps 1000 --noise_schedule linear --rescale_learned_sigmas False --rescale_timesteps False --lr 1e-4 --batch_size 8 -
For sampling, run:
python scripts/segmentation_sample.py --data_dir (where you put data folder)/data/testing --out_dir output data direction --model_path saved model --image_size 256 --num_channels 128 --class_cond False --num_res_blocks 2 --num_heads 1 --learn_sigma True --use_scale_shift_norm False --attention_resolutions 16 --diffusion_steps 1000 --noise_schedule linear --rescale_learned_sigmas False --rescale_timesteps False --num_ensemble 5
...
It is simple to run MedSegDiff on the other datasets. Just write another data loader file following ./guided_diffusion/isicloader.py or ./guided_diffusion/bratsloader.py. Welcome to open issues if you meet any problem. It would be appreciated if you could contribute your dataset extensions. Unlike natural images, medical images vary a lot depending on different tasks. Expanding the generalization of a method requires everyone's efforts.
To train a fine model, i.e., MedSegDiff-B in the paper, set the model hyperparameters as:
--image_size 256 --num_channels 128 --class_cond False --num_res_blocks 2 --num_heads 1 --learn_sigma True --use_scale_shift_norm False --attention_resolutions 16
diffusion hyperparameters as:
--diffusion_steps 1000 --noise_schedule linear --rescale_learned_sigmas False --rescale_timesteps False
To speed up the sampling:
--diffusion_steps 50 --dpm_solver True
run on multiple GPUs:
--multi-gpu 0,1,2 (for example)
training hyperparameters as:
--lr 5e-5 --batch_size 8
and set --num_ensemble 5 in sampling.
Run about 100,000 steps in training will be converged on most of the datasets. Note that although loss will not decrease in most of the later steps, the quality of the results are still improving. Such a process is also observed on the other DPM applications, like image generation. Hope someone smart can tell me why🥲.
I will soon publish its performance under smaller batch size (suitable to run on 24GB GPU) for the need of comparison🤗.
A setting to unleash all its potential is (MedSegDiff++):
--image_size 256 --num_channels 512 --class_cond False --num_res_blocks 12 --num_heads 8 --learn_sigma True --use_scale_shift_norm True --attention_resolutions 24
Then train it with batch size --batch_size 64 and sample it with ensemble number --num_ensemble 25.
Welcome to contribute to MedSegDiff. Any technique can improve the performance or speed up the algorithm is appreciated🙏. I am writting MedSegDiff V2, aiming at Nature journals/CVPR like publication. I'm glad to list the contributors as my co-authors🤗.
- Fix bugs in BRATS. Add BRATS example.
- Release REFUGE and DDIT dataloaders and examples
- Speed up sampling by DPM-solver
- Inference of depth
- Fix bugs in Multi-GPU parallel
- Sample and Vis in training
- Release pre processing and post processing
- Release evaluation
- Deploy on HuggingFace
- yaml configuration
Code copied a lot from openai/improved-diffusion, WuJunde/ MrPrism, WuJunde/ DiagnosisFirst, LuChengTHU/dpm-solver, JuliaWolleb/Diffusion-based-Segmentation, hojonathanho/diffusion, guided-diffusion, bigmb/Unet-Segmentation-Pytorch-Nest-of-Unets, nnUnet, lucidrains/vit-pytorch
Please cite
@article{wu2022medsegdiff,
title={MedSegDiff: Medical Image Segmentation with Diffusion Probabilistic Model},
author={Wu, Junde and Fang, Huihui and Zhang, Yu and Yang, Yehui and Xu, Yanwu},
journal={arXiv preprint arXiv:2211.00611},
year={2022}
}
@article{wu2023medsegdiff,
title={MedSegDiff-V2: Diffusion based Medical Image Segmentation with Transformer},
author={Wu, Junde and Fu, Rao and Fang, Huihui and Zhang, Yu and Xu, Yanwu},
journal={arXiv preprint arXiv:2301.11798},
year={2023}
}