- Documentation is now available at: https://med-imagetools.readthedocs.io
- Dependencies have been reduced for a lighter install.
torch
andtorchio
dependencies have been moved to an extra pip install flag. Usepip install med-imagetools[torch]
to use the Dataset feature and
- AutoPipeline CLI
nnunet
nnU-Net compatibility mode- Built-in train/test split for both normal/nnU-Net modes
random_state
for reproducible seeds- Region of interest (ROI) yaml dictionary intake for RTSTRUCT processing
- Markdown report output post-processing
continue_processing
flag to continue autopipelinedry_run
flag to only crawl the dataset
Med-Imagetools, a python package offers the perfect tool to transform messy medical dataset folders to deep learning ready format in few lines of code. It not only processes DICOMs consisting of different modalities (like CT, PET, RTDOSE and RTSTRUCTS), it also transforms them into deep learning ready subject based format taking the dependencies of these modalities into consideration.
A medical dataset, typically contains multiple different types of scans for a single patient in a single study. As seen in the figure below, the different scans containing DICOM of different modalities are interdependent on each other. For making effective machine learning models, one ought to take different modalities into account.
Fig.1 - Different network topology for different studies of different patientsMed-Imagetools is a unique tool, which focuses on subject based Machine learning. It crawls the dataset and makes a network by connecting different modalities present in the dataset. Based on the user defined modalities, med-imagetools, queries the graph and process the queried raw DICOMS. The processed DICOMS are saved as nrrds, which med-imagetools converts to torchio subject dataset and eventually torch dataloader for ML pipeline.
Fig.2 - Med-Imagetools AutoPipeline diagrampip install med-imagetools
conda create -n mit
conda activate mit
pip install med-imagetools
conda create -n mit
conda activate mit
pip install -e git+https://github.com/bhklab/med-imagetools.git
This will install the package in editable mode, so that the installed package will update when the code is changed.
Med-Imagetools takes two step approch to turn messy medical raw dataset to ML ready dataset.
- Autopipeline: Crawls the raw dataset, forms a network and performs graph query, based on the user defined modalities. The relevant DICOMS, get processed and saved as nrrds
autopipeline\ [INPUT DIRECTORY] \ [OUTPUT DIRECTORY] \ --modalities [str: CT,RTSTRUCT,PT] \ --spacing [Tuple: (int,int,int)]\ --n_jobs [int]\ --visualize [flag]\ --nnunet [flag]\ --train_size [float]\ --random_state [int]\ --roi_yaml_path [str]\ --continue_processing [flag]\ --dry_run [flag]
- class Dataset: This class converts processed nrrds to torchio subjects, which can be easily converted to torch dataset
from imgtools.io import Dataset subjects = Dataset.load_from_nrrd(output_directory, ignore_multi=True) data_set = tio.SubjectsDataset(subjects) data_loader = torch.utils.data.DataLoader(data_set, batch_size=4, shuffle=True, num_workers=4)
These google collab notebooks will introduce the main functionalities of med-imagetools. More information can be found here
Thanks to the following people who have contributed to this project:
If you have any questions/concerns, you can reach the following contributors at sejin.kim@uhnresearch.ca
This project uses the following license: MIT License