wenyi-li / FairDiff

The official code repository for paper "FairDiff: Enhancing Fairness in Fundus Image Segmentation with Point-Image Diffusion".

FairDiff: Enhancing Fairness in Fundus Image Segmentation with Point-Image Diffusion

The official code repository for paper "FairDiff: Enhancing Fairness in Fundus Image Segmentation with Point-Image Diffusion".

Installation

You can install the necessary dependencies using the provided conda environment YAML file:

# Create the environment
conda env create -f env.yml
# Activate the environment
conda activate FairDiff

Data Preparation

Please follow the instruction in this link to download Harvard-FairSeg Dataset in .npz format.

Convert the original fundus data ( .npz format) to images and labels ( .png format). Ensure to modify the data path accordingly:

python data_convert.py 

Point-Mask Generation

Data preparation

Note that there are several categories: 'ethnicity', 'gender', 'language', 'maritalstatus', and 'race'. Each category contains different attributes. For example, under the gender category, there are two attributes: Female and Male.

cd PointMaskGen/data/pointcloud
python labeloutline2pc.py --category_path /datasets/MedicalImage/txt/gender
# --category_path saves the image paths of all the attributes under a category.
python generate-txt.py --categories gender

Training

cd Point-Image/Pointdiffusion
python train.py --categories gender --attributes Female

Mask Generation

python inference.py --categories gender --attributes Female --resume_path xxx.ckpt

Mask-Image Generation

We utilize ControlNet to generate images from masks.

Training

Train a mask-image generation model for each group with different sensitive attributes to generate corresponding images:

cd MaskImageGen
python train.py --attr_type <attr> --name <specific attribute to train>

Images Generation

Generate corresponding images using:

CUDA_VISIBLE_DEVICES=0 python infer.py  \
    --attr_type ${ATTR_TYPE} --name ${NAME} \
    --ckpt ${CHECKPOINT} \
    --images 10000 \
    --save_path ${SAVE_PATH}

or

bash infer.sh

Data packing

After generating images, pack images and labels into .npz files for the next segmentation task:

python packing.py

.npz files contain the following keys:

• slo_fundus: Scanning laser ophthalmoscopy (SLO) fundus image

• disc_cup_mask: Disc and cup masks for the corresponding SLO fundus image

• age: Patient's age

• gender: 0 - Female, 1 - Male

• race: 0 - Asian, 1 - Black, 2 - White

• ethnicity: 0 - Non-Hispanic, 1 - Hispanic, -1 - Unknown

• language: 0 - English, 1 - Spanish, 2 - Others, -1 - Unknown

• maritalstatus: 0 - Married or Partnered, 1 - Single, 2 - Divorced, 3 - Widowed, 4 - Legally Separated, -1 - Unknown

Evaluate

To assess the generation quality, employ FID, minimum matching distance (MMD), and coverage score (COV):

Calculate FID:

python -m pytorch_fid path/to/dataset1 path/to/dataset2

Calculate MMD and COV:

python evaluate.py --ref_folder original_image_path --gen_folder generate_image_path --csv_path save_path

Fair Segmentation

To verify the impact of our synthetic data on segmentation and fairness, we selected two segmentation models, including a classic model TransUNet and a larger model SAMed.

# Using TransUNet
cd Segmentation/TransUNet

# Using SAMed
cd Segmentation/SAMed

Data Combination

In order to ensure that the data used in all experiments is strictly consistent, the mixing methods of different types of data and the usage of synthetic data under different conditions have been written into txt files and stored in the lists/ directory.

You can use the command below for more data list generation.

python data_combination.py

In order to visually demonstrate our experimental process, the list files generated in our environment have not been deleted, but have been anonymized.

Training

First, prepare the dataset according to the synthesis and balancing methods mentioned in the paper, and use the generated list mentioned above for the experiment you want to complete.

In train.sh, please specify the base_dir, list_dir and attributes you want to use. Finally, run this command.

./train.sh

Testing

For testing, please specify the base_dir, attribute, path of pretrained lora checkpoint, and output_dir. Then run this command.

./test.sh