Med-Imagetools: Transparent and Reproducible Medical Image Processing Pipelines in Python

GitHub repo size GitHub contributors

Latest Updates (v1.2.0) - Feb 5th, 2024

Documentation is now available at: https://med-imagetools.readthedocs.io
Dependencies have been reduced for a lighter install. torch and torchio dependencies have been moved to an extra pip install flag. Use pip install med-imagetools[torch] to use the Dataset feature and

Med-ImageTools core features

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 autopipeline
dry_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.

Introduction

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 patients

Med-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 diagram

Installing med-imagetools

pip install med-imagetools

(recommended) Create new conda virtual environment

conda create -n mit
conda activate mit
pip install med-imagetools

(optional) Install in development mode

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.

Getting Started

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)

Demo (Outdated as of v0.4)

These google collab notebooks will introduce the main functionalities of med-imagetools. More information can be found here

Tutorial 1: Forming Dataset with med-imagetools Autopipeline

Tutorial 2: Machine Learning with med-imagetools and torchio

Contributors

Thanks to the following people who have contributed to this project:

Contact

If you have any questions/concerns, you can reach the following contributors at sejin.kim@uhnresearch.ca

License

This project uses the following license: MIT License

bhklab / med-imagetools

readme