A Comprehensive Federated Learning Framework for Diabetic Retinopathy Grading and Lesion Segmentation

Getting started

1. Create conda environment and install dependencies:

conda create -y -n TJDR-FL python=3.8

conda activate TJDR-FL

conda install -y pytorch==1.10.2 torchvision==0.11.3 torchaudio==0.10.2 cudatoolkit=11.3 -c pytorch

pip install pip -U
pip install -r requirements.txt

2. Prepare dataset:

dataset_dir="../TJDR-FL/task/datas/"

• For IDRiD, DDR-seg, DDR-cls, and APTOS2019 datasets, download the official dataset into the corresponding dir and subsequently run ../TJDR-FL/task/experiments/{dataset_name}.py

• For TJDR, download the dataset from the link our provided and move all files into ../TJDR-FL/task/datas/TJDR

Run the code and Training

Code runs according to *.yaml config file, we provide two methods to run the code:

1. Run from base_config.yaml

Training form configs/base_config.yaml as follows:

cd task
python run.py  

Note: Please make sure configs/base_config.yaml is prepared as you expected, as the code will be executed based on it in the case.

2. Run from base config templates

For convenience, we provide base config templates to perform training, in which code run from configs/base_configs/{classification|segmentation}/{dataset}.yaml. You can copy the template to override the configs/base_config.yaml to run the code.

Args:

-b, --base_config_path : Specified the path of base config, default is base_config.yaml  , Optional
-g, --gpu              : Specified gpu to run, default is specified by config files      , Optional
-n, --network          : Enable Network to parallel computing, default is false
--all_gpu              : Enable all gpu to parallel computing, default is false
--host                 : Cloud host, default is initialized by 'configs/base_config.yaml', Optional
--port                 : Cloud port, default is initialized by 'configs/base_config.yaml', Optional

Our TJDR dataset is available at here.

System information

We train our model on 6 NVIDIA GeForce RTX 3090 GPUs with a 24GB memory per-card. Testing is conducted on the same machines.