This is the implementation of the MICCAI 2021 Paper "I-SECRET: Importance-guided fundus image enhancement via semi-supervised contrastive constraining".

New work LED is public, which also includes the pre-trained weights of I-SECRET. 03/09/2023

1. Firstly, download EyeQ dataset from EyeQ.
2. Split the dataset into train/val/test according to the EyePACS challenge.
3. Run

python tools/degrade_eyeq.py --degrade_dir ${DATA_PATH}$ --output_dir $OUTPUT_PATH$ --mask_dir ${MASK_PATH}$ --gt_dir ${GT_PATH}$.

Note that this scipt should be applied for usable dataset for cropping pre-processing.

4. Make the architecture of the EyeQ directory as:

.
├── 
├── train
│   └── crop_good
│   └── degrade_good
│   └── crop_usable
├── val
│   └── crop_good
│   └── degrade_good
│   └── crop_usable
├── test
│   └── crop_good
│   └── degrade_good
│   └── crop_usable

Here, the crop_good is the ${GT_PATH}$ in the step 3, and degrade_good is the ${OUTPUT_PATH}$ in the step 3.

Run

pip install -r requirements.txt

Run the baseline model

python main.py --model i-secret --lambda_rec 1 --lambda_gan 1 --data_root_dir ${DATA_DIR}$ --gpu ${GPU_INDEXS}$ -- batch size {BATCH_SIZE}$  --name baseline --experiment_root_dir ${LOG_DIR}$

Run the model with IS-loss

python main.py --model i-secret --lambda_is 1 --lambda_gan 1 --data_root_dir ${DATA_DIR}$ --gpu ${GPU_INDEXS}$ -- batch size {BATCH_SIZE}$  --name is_loss --experiment_root_dir ${LOG_DIR}$

Run the I-SECRET model

python main.py --model i-secret --lambda_is 1 --lambda_icc 1 --lambda_gan 1 --data_root_dir ${DATA_DIR}$ --gpu ${GPU_INDEXS}$ --batch_size {BATCH_SIZE}$  --name i-secret --experiment_root_dir ${LOG_DIR}$

Go to the ${LOG_DIR}$ / ${EXPERIMENT_NAME}$ / checkpoint, run

tensorboard --logdir ./ --port ${PORT}$

then go to localhost:${PORT}$ for detailed logging and visualization.

Run

python main.py --test --resume 0 --test_dir ${INPUT_PATH}$ --output_dir ${OUTPUT_PATH}$ --name ${EXPERIMENT_NAME}$ --gpu ${GPU_INDEXS}$ -- batch size {BATCH_SIZE}$ 

Please note that the metric outputted by test script is under the PyTorch pre-process (resize etc.). It is not precise. Therefore, we need to run the evaluation scipt for further evaluation.

python tools/evaluate.py --test_dir ${OUTPUT_PATH}$ --gt_dir ${GT_PATH}$

We apply the iter-Net framework. We simply replace the test set with the degraded images/enhanced images. For more details, please follow IterNet.

• Cleaning codes
• More SOTA backbones (ResNest ...)
• WGAN loss
• Internal evaluations for down-sampling tasks

Thanks for CutGAN for the implementation of patch NCE loss, EyeQ_Enhancement for degradation codes, Slowfast for the distributed training codes