Scripts

We provide several scripts for batch processing of datasets and performance evaluation.

Common Datasets

A common dataset should have a dataset_path that contains all sub directories. Files in different sub directories may have different suffix. These parameters may need to be specified in the following scripts.

Extract Files

Code refers to extract_files. This script copies or moves files from a source directory to target directory based on a template directories. Files from source directories should have the same names with the template directory.

python extract_files.py --source_dir <source_dir> --template_dir <template_dir=.> --output_dir <output_dir=> --suffix <suffix=\'\'> --method <method=copy>

Crop Image

Code refers to crop_by. This script uses bounding box to crop images.

If you run the command with --separately option, we compute a the boundingbox for each sample based on the label or raw data (specified by crop_by). The boundingbox is computed for non-background region.

If you run the command without --separately option, you can choose to specify a boundingbox, or the boundingbox will be computed based on all samples. Note that all samples should have the same shape in this situation.

python crop_by.py -p <dataset_path> --sub_dirs <sub_dirs=.> --suffix <suffix=\'\'> --crop_by <crop_by=None> -o <output_dir=.> -m <margin = 20,20,20>  -b <background = 0.0> --boundingbox <boundingbox = None>

Transfer h5 to images (png)

Code refers to h5_to_png. H5 is a dict-like file, which can store multiple keys and its corresponding contents. To transfer H5 file to png, you need to specify the path and sub directories of h5 dataset, the keys corresponding to 2D/3D images, suffix of h5 files, and output directory. Extracted image will be stored in the sub directories of the output directory.

python h5_to_png.py -p <dataset_path> -o <output_dir=.> --sub_dirs <sub_dirs=\'.\'> --suffix <suffix=.h5> --keys <keys=\'\'> '

Randomly Split a Dataset

Code refers to random_split_k_fold. This script randomly split a dataset into k folds. Each fold has a same file structure with original datasets.

python random_split_k_fold.py -p <dataset_path> --sub_dirs <sub_dirs=.> --suffix <suffix=.> -o <output_dir=.> -k <kfold=5> --method <method=copy>

FastMRI Dataset

We use FastMRI to evaluate the performance of reconstruction and generation. Specially, we construct noisy image by masking out 90% k-space content and reconstructing through zero-filled algorithm. Code refers to fastmri_masked_zero_fill.

python fastmri_masked_zero_fill --data_path <data_path> --output_path <output_path> --challenge <singlecoil or multicoil>

Then we can use h5_to_png script to transfer the h5 file into png images.

Performance Evaluation

We also provide a script for performance evaluation. Code refers to test_time_and_space. In specific, we construct models based on the model_config.yaml, and create random tensors to stimulate the forward and backward processing. Time and space usage are recorded and printed after the test. To run this script, we don’t need to specify any parameter in the command.

python test_time_and_space.py

An example of evaluation results refers to eval_time_and_space.log.

Select Samples for Visualization

Sometimes we tried different methods like A, B, and C. And we might get a evaluation results like the follows:

\[\text{ACC}(A) > \text{ACC}(B) > \text{ACC}(C)\]

However, the accuracy is computed over the whole dataset. If we randomly choose a sample from the datasets, the results might not follow the above relation. Also, we may want to select the sample with enough foreground regions for visualization.

This script can select samples whose predictions follow some specified relations and filter out those samples that don’t contains a mininum number of foreground pixels. You can also choose to save the colorized segmented results. This is particularly helpful when you need to make a figure to compare the results of different methods in your paper!!

python select_samples_and_colorize.py --result_path path/to/predictions \
# sub dirs of different methods
--sub_dirs ResNet,ResNet_HSeg,UNet,UNet_HSeg,UNet_Atten,UNet_Atten_HSeg,SwinU,SwinU_HSeg,MambaU,MambaU_HSeg \
--suffix .png \
--target_path path/to/results/seg/CVC-ClinicDB/target \
# sub dirs of targets. because the targets might have different size.
# if the methods have the same targets, just setting the target_path is ok
--target_sub_dirs ResNet,ResNet_HSeg,UNet,UNet_HSeg,UNet_Atten,UNet_Atten_HSeg,SwinU,SwinU_HSeg,MambaU,MambaU_HSeg  \
--target_suffix _tar.png \
# set input path for visualization.
--input_path path/to/input/CVC-ClinicDB/raw \
# sub dirs of inputs (raw images). because the inputs might have different size.
# if the methods have the same inputs, just setting the input_path is ok
--input_subdir ResNet_HSeg,UNet,UNet_HSeg,UNet_Atten,UNet_Atten_HSeg,SwinU,SwinU_HSeg,MambaU,MambaU_HSeg\
--input_suffix
# output_dir, where the selected and visualized results will be saved
--output_dir path/to/output
# conditions are the specified relations
--conditions '0<1,2>3,4>5,6>7,8>9' \
# choose a evaluation metric to compare
--eval Dice
# if this optional is set, we choose the middle slice to represent a 3D image.
# otherwise, we will search all slices (might take a much longer time).
--compute_middle_for_3d
# for segmentation, the ratio of foreground points or pixels should be larger than 0.01
--minimum_ratio 0.01
# the differences among scores should be larger than 0.05
--score_margin 0.05

Render Points and Mesh

Render point cloud or mesh with mitsuba. The path of input, transformation, color, illumination, and other rendering-related parameters can be adjusted in the codes. The render results are saved in mitsuba_scene.png.