DiffuseExpand (expanding dataset for 2D medical image segmentation using diffusion models)
Getting Started
Build Environment
First, download our repo and then enter our repo
cd DiffuseExpand
For environmental establishment, we include .yaml files.
If you have an RTX 30XX GPU (or newer), run
conda env create -f requirements_11_3.yaml
If you have an RTX 20XX GPU (or older), run
conda env create -f requirements_10_2.yaml
You can then activate your conda environment with
conda activate diffuseexpand
Download pre-training checkpoints
wget https://openaipublic.blob.core.windows.net/diffusion/jul-2021/256x256_diffusion.pt
Download Datasets
The datasets used in our work are COVID-19 and CGMH Pelvis, please download them from the website and remember the path to the corresponding dataset:
Fine-tune diffusion model (Stage I)
We used 8 Tesla A100 GPU for the experiment, with a batchsize of 2 on each Tesla A100 GPU. For COVID-19, we run
mkdir ./stage2
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python stage1_train.py --dataset COVID19 \
--data_path /path/to/covid-chestxray-dataset/image --csv_path /path/to/covid-chestxray-dataset/metadata.csv \
--save_path ./stage2 --unet_ckpt_path ./256x256_diffusion.pt --cuda_devices 0,1,2,3,4,5,6,7
After that, we can get ./stage2/model_stage2_covid_30000.pt.
For CGMH Pelvis, the first step is to change the name of the checkpoint to model_stage2_cgmh_{self.resume_step + self.step}.pt in utils/train_utils.py:line 219. Then, we run
mkdir ./stage2
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python stage1_train.py --dataset CGMH \
--data_path /path/to/CGMH_PelvisSegment \
--save_path ./stage2 --unet_ckpt_path ./256x256_diffusion.pt --cuda_devices 0,1,2,3,4,5,6,7
After that, we can get ./stage2/model_cgmh_covid_30000.pt.
Train Segmenter (Stage II)
We used 2 Tesla A100 GPU for the experiment, with a batchsize of 8 on each Tesla A100 GPU. For COVID-19, we run
mkdir ./stage3
CUDA_VISIBLE_DEVICES=0,1 python stage2_train.py --dataset COVID19 \
--data_path /path/to/covid-chestxray-dataset/image --csv_path /path/to/covid-chestxray-dataset/metadata.csv \
--save_path ./stage3 --cuda_devices 0,1
After that, we can get ./stage3/stage3_covid19_model_10000.pt.
For CGMH Pelvis, the first step is to change the name of the checkpoint to stage3_cgmh_model_{step}.pt in stage2_train.py:line 355. Then, we run
mkdir ./stage3
CUDA_VISIBLE_DEVICES=0,1 python stage2_train.py --dataset CGMH \
--data_path /path/to/CGMH_PelvisSegment \
--save_path ./stage3 --cuda_devices 0,1
After that, we can get ./stage3/stage3_cgmh_model_10000.pt.
Synthesize Image-Mask pairs (Stage III)
For synthesizing Image-Mask pairs with COVID-19, we run
mkdir ./stage3_covid19
python stage3_covid_test.py --save_path ./stage3_covid19 --dpm-checkpoint ./stage2/model_stage2_covid_30000.pt \
--cls-checkpoint ./stage3/stage3_covid19_model_10000.pt --synthesize-number 500
After that, we can get synthesized sample pairs in the folder ./stage3_covid19.
For synthesizing Image-Mask pairs with CGMH Pelvis, we run
mkdir ./stage3_cgmh
python stage3_cgmh_test.py --save_path ./stage3_cgmh --dpm-checkpoint ./stage2/model_stage2_cgmh_30000.pt \
--cls-checkpoint ./stage3/stage3_cgmh_model_10000.pt --synthesize-number 500
After that, we can get synthesized sample pairs in the folder ./stage3_cgmh.
Choose high quality Image-Mask pairs (Stage IV)
Before proceeding to Stage IV, two additional things need to be done: first, train a unet using eval.py and save its corresponding checkpoint, and second, synthesize enough samples pairs at Stage III to facilitate the selection of high-quality sample pairs.
Then for COVID-19, we can run
mkdir ./stage4_covid19
python stage4_train.py --unet-checkpoint /path/to/unet/checkpoint --stage3-output ./stage3_covid19 \
--stage4-output ./stage4_covid19
And for CGMH Pelvis, we can run
mkdir ./stage4_cgmh
python stage4_train.py --unet-checkpoint /path/to/unet/checkpoint --stage3-output ./stage3_cgmh \
--stage4-output ./stage4_cgmh
Train the validated model
For COVID-19, we can run
python eval.py --dataset=COVID19 --loss_type sigmoid_l1 --model=Unet --train_epochs=50 \
--generate_data_path ./stage4_covid19 \
--data_path /path/to/covid-chestxray-dataset/image --csv_path /path/to/covid-chestxray-dataset/metadata.csv \
And for CGMH Pelvis, we can run
python eval.py --dataset=CGMH --loss_type sigmoid_l1 --model=Unet --train_epochs=50 \
--generate_data_path ./stage4_cgmh \
--data_path /path/to/CGMH_PelvisSegment \