Test-time Bi-directional Adaptation between Image and Model for Robust Segmentation

In this paper, we aim to develop a general way to generalize existing segmentation models to samples with unknown appearance shift when deployed in daily clinical practice. We propose an effective test-time bi-directional adaptation solution for this aim by combining two complementary strategies.

Prerequisites

• Python 3.8
• Check the requirements.txt

pip install -r requirements.txt

Modify data path to your own in config.py

cfg = Config({
    "train_dir": "",  ## Path to image folder
    "label_dir": "",  ## Path to label folder
})

Train

CUDA_VISIBLE_DEVICES=<gpu_id> python code/train.py.py --dataset "your dateset name"

Test (style transfer)

Arguments adain, osa, wosa for different style transfer modules

python code/seg_test.py --wosa
python code/seg_test.py --osa
python code/seg_test.py --adain

Test (Augmented self-supervision)

python code/seg_augself.py  ## AugSelf
python code/seg_augself.py --wosa  ## WOSA-AugSelf