tta_readme.md

Test-Time Adaptation Module

Test-time adaptation module, using a pretrained model or provided segmentations as a basis and finetune on your image data.

prediction with pre-trained model

Prepare prerequisite model

We currently provide 3 pre-trained models, you can download them to make prediction on your images. The models are ...

To access them, make a directory for the pre-trained models within the vessel_code folder:

mkdir ./pretrained_models/

Download the pre-trained model from osf:

osf -p abk4p fetch /osfstorage/pretrained_models/manual_0429 ./pretrained_models/manual_0429

Test-time adpatation without providing proxy segmentation

You could apply this module directly on your on data without providing the proxy segmentation. This module will automatically generate proxies and finetune the model. If you set prep_mode to 4, which means no preprocessing will happen, then you don't have to set a path to store the preprocessed images. Note that the input path must only contain the nifti images for processing and be different to the output path.

python test_time_adaptation.py --ds_path $path_to_images --out_path $path_to_output --pretrained $path_to_pretrained_model --prep_mode 4 --ep $n_epochs --lr 1e-3 

If you set prep_mode to 1,2 or 3, which means (1) N4 bias field correction, (2)denosing, or (3) both N4 biasfield correction and denoising will be applied to the input images, then you have to set a path to store the preprocessed images. In the following example, we set the preprocessing mode to "applying N4 bias field correction only (mode 1)":

python test_time_adaptation.py --ds_path $path_to_images --out_path $path_to_output --ps_path $path_to_preprocessed_images --pretrained $path_to_pretrained_model --prep_mode 1 --ep $n_epochs --lr 1e-3 

Test-time adpatation with provided proxy segmentation

You can provide a proxy segmentation which will be used for the test-time-adaption instead of the automatically generated proxies. This segmentation can be created using any methods, for example manual labelling or other, non-deep-learning methods. We will use the filenames to match the raw images and proxy segmentations. If you set prep_mode to 4, which means no preprocessing will happen, then you don't have to set a path to store the preprocessed images.

python test_time_adaptation.py --ds_path $path_to_images --px_path $path_to_proxy_labels --out_path $path_to_output --pretrained $path_to_pretrained_model --prep_mode 4 --ep $n_epochs  --lr 1e-3

If you set prep_mode to 1,2 or 3, which means (1) N4 bias field correction, (2)denosing, or (3) both N4 biasfield correction and denoising will happen, then you have to set a path to store the preprocessed images. In the following example, we set the preprocessing mode to "applying N4 bias field correction only (mode 1)":

python test_time_adaptation.py --ds_path $path_to_images --px_path $path_to_proxy_labels --out_path $path_to_output --ps_path $path_to_preprocessed_images --pretrained $path_to_pretrained_model --prep_mode 1 --ep $n_epochs --lr 1e-3