GlaucoDiff

🛠️ Install dependencies

conda create -n GlaucoDiff python=3.11.7
conda activate GlaucoDiff
pip install -r requirements.txt

📁 Data preparation

The Harvard-FairSeg dataset (named as 10k) can be accessed via this link, and the Harvard-FairVLMed dataset (named as fairvlmed10k) can be accessed via this link.

We provide the data_summary.csv and filter_file.txt files for both datasets, which contain the filenames used in our experiments, along with information on whether each file is used for the training, validation, or test set. In addition to the demographic information and medical records provided by the source data, we categorize ages under 65 as "young" and ages 65 and above as "elderly".

Move data files to your own directory path:

mv data/fairvlmed10k/* /src/fairvlmed10k
mv data/10k/* /src/10k

Download the datasets 10k and fairvlmed10k to GlaucoDiff/data. The directory should look like

10k
.
|-- Test 
│   ├── data_08001.npz
│   ├── data_08002.npz
|-- Training 
│   ├── data_00001.npz
│   ├── data_00002.npz
|-- Validation 
│   ├── data_07515.npz
│   ├── data_07516.npz
|-- data_summary.csv
|-- filter_file.txt

fairvlmed10k
.
|-- Test 
│   ├── data_08001.npz
│   ├── data_08002.npz
|-- Training 
│   ├── data_00001.npz
│   ├── data_00002.npz
|-- Validation 
│   ├── data_07001.npz
│   ├── data_07002.npz
|-- data_summary.csv
|-- filter_file.txt

🚀 Train

First

The complete code for generating glaucoma images will be made publicly available after the paper is accepted. Coming Soon.

Second

Here, we provide only the training code and pre-trained models for TransUNet. The code for SAM, Unet, and the complete sample selection process will be made publicly available after the paper is published. Coming Soon.

Trained segmentation model:

cd src/TransUNet-main
python train.py --root_path /root/ --dataset 10k --vit_name R50-ViT-B_16

Download pre-trained segmentation model and place it in the specified directory:

Model Name	Download Link	Description
TU_pretrain_R50-ViT-B_16_skip3_epo150_bs24_224	Google Drive	Place the pre-trained segmentation model under the directory scr/TransUNet-main/model

Use segmentation models to grade the generated images.

python test.py --is_savenii --root_path /sample/gen_image/10k --dataset sd_gen0 --vit_name R50-ViT-B_16
or
python test.py --is_savenii --root_path /sample/gen_image/fairvlmed10k --dataset sd_gen0 --vit_name R50-ViT-B_16

After running the segmentation model, you will obtain the file sd_gen_metric{seed}.csv

Third

Organize the generated data to form the following structure:

data_root
.
|-- 10k 
│   ├── Test
│   ├── Training
│   ├── Validation
│   ├── data_summary.csv
│   ├── filter_file.txt
|-- fairvlmed10k 
│   ├── Test
│   ├── Training
│   ├── Validation
│   ├── data_summary.csv
│   ├── filter_file.txt
|-- gen_image 
│   ├── 10k
│       ├── sd_gen0
│           ├── data_00008_d3ab.png
│           ├── data_00008_d3ab_generate.png
│       ├── sd_gen_metric.csv
│   ├── fairvlmed10k
│       ├── sd_gen0
│           ├── data_00003_6a47.png
│           ├── data_00003_6a47_generate.png
│       ├── sd_gen_metric.csv

Forth

Training the classification model

cd src/FundusProcessModel
python Train.py --pretrain_model  --model efficientnet-b0 --data_root /root/ --dataset fairvlmed10k --use_fake_data  --epochs 100 --top_precentege 0

Common parameters:

• --model: Model selection
• --epochs: Number of epochs
• --device: Training device (cuda:0/cpu)
• --data_root: Root directory of the dataset
• --dataset: Name of the dataset used
• --pretrain_model: Whether to use a pre-trained model
• --use_fake_data: Whether to use generated data
• --top_percentage: Percentage of generated data to use
• --balance_data: Whether to balance the quantities of each class
• --balance_attribute: Attribute to balance
• --best_model_path: Path to the loaded model (useful only for running the test set)

🧪 only Test

Download pre-trained models and place it in the specified directory:

Model	Download Link	Description
efficientnet-b0 10k_best_auc	Google Drive	Trained by the 10k training set with generated samples
efficientnet-b0 fairvlmed10k_best_auc	Google Drive	Trained by the fairvlmed10k training set with generated samples

If only testing the model performance, please run the code below:

python --model efficientnet-b0 --data_root /root/ --dataset fairvlmed10k --best_model_path checkpoints/label_classification/wt5ic8g1/best_auc.pth