Prepare Data¶
For all methods, SimpleITK is used as the image IO and our test is based on “nii.gz” format.
Training for registration tasks¶
We assume there is three level folder, output_root_path/ data_task_name/ task_name. The output_root_path/data_task_name must be created and well organized as following before running the train script.
- In data_task_name, each folder refer to different preprocessing strategies or different datasets, i.e. lung tasks and brain tasks. It might as well could be different splits of the data too.
- In task_name, each folder refer to a specific setting, for a specific dataset setting.
Thus, the folders will be created as output_root_path/data_task_name/task_name.
For training tasks, the data should be organized as following:
- train, val, test, debug (subset of train data, to check overfit) folder should be put under output_root_path/data_task_name, each of the folder should include pair_path_list.txt and pair_name_list.txt
- pair_path_list.txt: each line of the txt include 4 terms: source_path, target_path, labels_for_source_image_path, labels_for_target_image_path
- pair_name_list.txt: each line of the txt include 1 term: the pair name. The file has line to line corresponded with pair_path_list.txt
Alternatively, prepare_data.py script is capable of generating corresponding directory structure, just using it will be enough.
Inference/Optimization in registration tasks¶
For inference (with pretrained model) or optimization tasks, we support two kind of input format.
- put image list in command line.
- put image list in a .txt file, each line of the txt include 4 terms: s_pth t_pth ls_path lt_path
s refers to source, t refers to target, ls refers to label of source (leave empty or string “None” if not exist), lt refers to label of target (leave empty or string “None” if not exist)*
Example:
- command line:
python demo_for_easyreg_eval.py --setting_folder_path ./demo_settings/mermaid/opt_vsvf --gpu_id 0 --task_output_path ./demo_output/mermaid/opt_vsvf -s ./oai_examples/9352883_20051123_SAG_3D_DESS_LEFT_016610798103_image.nii.gz ./oai_examples/9761431_20051103_SAG_3D_DESS_RIGHT_016610945809_image.nii.gz -t ./oai_examples/9403165_20060316_SAG_3D_DESS_LEFT_016610900302_image.nii.gz ./oai_examples/9211869_20050131_SAG_3D_DESS_RIGHT_016610167512_image.nii.gz
- txt file:
python demo_for_easyreg_eval.py --setting_folder_path ./demo_settings/mermaid/opt_vsvf --gpu_id 0 --task_output_path ./demo_output/mermaid/opt_vsvf --pair_txt_path ./oai_examples.txt
Train for segmentation tasks¶
We assume there is three level folder, output_root_path/ data_task_name/ task_name. Our prepare_data.py script is capable of generating corresponding directory structure, just using it will be enough. The output_root_path/data_task_name must be created and well organized as following before running the train script.
- In data_task_name, each folder refer to different preprocessing strategies or different datasets, i.e. lung tasks and brain tasks. It might as well could be different splits of the data too.
- In task_name, each folder refer to a specific setting, for a specific dataset setting.
Thus, the folders will be created as output_root_path/data_task_name/task_name.
For training tasks, the data should be organized as following:
- train, val, test, debug (subset of train data, to check overfit) folder should be put under output_root_path/data_task_name, each of the folder should include file_path_list.txt and file_name_list.txt
- file_path_list.txt: each line of the txt include 2 terms: img_path, labels_path
- file_name_list.txt: each line of the txt include 1 term: the filename. The file has line to line corresponded with file_path_list.txt
Inference in segmentation Tasks¶
For inference (with pretrained model) or optimization tasks, we support two kind of input format.
- put image list in command line.
- put image list in a .txt file, each line of the txt include 4 terms: s_pth t_pth ls_path lt_path
s refers to source, t refers to target, ls refers to label of source (leave empty or string “None” if not exist), lt refers to label of target (leave empty or string “None” if not exist)*
Example:
- command line:
python demo_for_seg_eval.py --setting_folder_path ./demo_settings/seg/lpba_seg_eval --gpu_id 0 --task_output_path ./demo_output/seg/lpba_seg_eval -i ./lpba_examples/s3.nii.gz ./lpba_examples/s27.nii.gz
- txt file:
python demo_for_seg_eval.py --setting_folder_path ./demo_settings/seg/lpba_seg_eval --gpu_id 0 --task_output_path ./demo_output/seg/lpba_seg_eval --file_txt_path ./lpba_examples.txt
Data Augmentation Tasks¶
We support two different data augmentation strategy, random augmentation and anatomical augmentation.
- For the random augmentation, we support Bspine augmentation and fluid-based random augmentation.
- For the anatomical augmentation, we support random sampling and data inter-/extra-polation.
Both tasks take a txt file recording file paths as input, items in the same line are separated by the space:
- For the random augmentation, the augmentation takes place among different lines, each line refers to a image and corresponding label (string “None” if not exist).
- For the anatomical augmentation, the augmentation takes place in a line, each line refers to a path of source image, paths of target images and the source label (string “None” if not exist), the labels of target images(None if not exist).
Additionally, an optional input is a txt file recording filename:
- For the random augmentation, each line include a image name.
- For the anatomical augmentation, each line include a source name and a series of target names.