SAUNA

SAUNA: Image-level Regression for Uncertainty-aware Retinal Image Segmentation

Dependencies

We provide file env.yaml for dependencies.

Installation

conda env create -f env.yaml
conda activate sauna
pip install -e .

Split data

cd mlpipeline/utils
python split_fives.py

SAUNA transform

cd mlpipeline/utils
python ./generate_uncertainty_masks.py --root <ROOT_DIR> --in_dir <GT_DIR>

The labels are available at URL.

Training

python -m mlpipeline.train.run experiment=${EXP_NAME} \
        model.params.cfg.arch=${ARCH_NAME}

where

• EXP_NAME: experiment setting can be fives_uncertainty_sem_seg (ours), fives_patch_sem_seg (for high-resolution-based methods), or fives_whole_sem_seg (for low-resolution-based methods).
• ARCH_NAME: architecture name can be Unet, UnetPlusPlus, IterNet, CTF-Net, CE-Net, DUnet, FR-Unet, DA-Net, or Swin-Unet.

Evaluation

python -m mlpipeline.train.evaluate \
    --config=${EXP_NAME} \
    --output_dir=/path/to/inference_results/${EXP_NAME} \
    --log_dir=/path/to/eval/${EXP_NAME} \
    --visual_dir=/path/to/visuals \
    --metadata_path=/path/to/test_split.pkl \
    --dataset_name=${DATASET} \
    --seeds=${SEEDS} \
    --folds=0,1,2,3,4

where

• DATASET: is either FIVES, STARE, DRIVE, CHASEDB1, or HRF.