Note: This repository is currently work in progress. Please contact us for more details here.
A Python wrapper for OpenJPH to enable encoding and/or decoding High-Throughput JPEG 2000 (HTJ2K) images.
In short, OpenJPH is an open-source implementation of the HTJ2K codec (JPEG2000 Part 15, ISO/IEC 15444-15, and ITU-T T.814), supporting features defined in JPEG 2000 Part 1. The code is written in C++ with color and wavelet transform steps taking advantage of SIMD instructions on Intel platforms. Unfortunately, this restricts encoding and/or decoding on image data directly from other commonly used programming languages. For example, another project openjphjs, developed by Chris Hafey, brings OpenJPH's capabilities to Javascript. Similarly, we aim to bring these capabilities to Python with direct support for HTJ2K encoding/decoding using openjphpy.
We use openjphpy in our implementation of the Medical Image Streaming Toolkit (MIST), an open-source toolkit to operationalize and democratize progressive resolution for large-scale medical imaging data infrastructures to accelerate data transmission and AI modelling. You can read our paper on MIST here.
For more resources regarding HTJ2K, please refer to:
- HTJ2K White Paper
- High throughput JPEG 2000 (HTJ2K): Algorithm, performance and potential
- High throughput block coding in the HTJ2K compression standard
If you use openjphpy in your publication, please cite the following:
- GitHub
@software{openjphpy,
author = {Kulkarni, Pranav},
title = {{openjphpy}},
month = {July},
year = {2023},
url = {https://github.com/UM2ii/openjphpy}
}
- Publication
@article{kulkarni2023one,
title={One Copy Is All You Need: Resource-Efficient Streaming of Medical Imaging Data at Scale},
author={Kulkarni, Pranav and Kanhere, Adway and Siegel, Eliot and Yi, Paul H and Parekh, Vishwa S},
journal={arXiv preprint arXiv:2307.00438},
year={2023}
}
Currently, openjphpy can be installed from pip or installed manually from source.
pip install openjphpy
git clone --recurse-submodules https://github.com/UM2ii/openjphpy
pip install -e openjphpy/
Documentation for openjphpy is available here.
Note: openjphpy is only supported on Linux based environments with support for other environments coming in the future.
We have provided an example notebook in this repository, along with 10 test images, to experiment with. You can find the example notebook here.
We also provide 15 sample medical images across X-ray, MRI, and CT modalities (5 images per modality), with data sourced from the NIH Chest X-Ray 14 and Medical Segmentation Decathlon (MSD) datasets. For portability, all data is stored as npy files.
Note: CT data is stored as uint16 and not in Hounsfield units.
Currently, openjphpy does not support encoding signed data and can only encode imaging data with types 8-bit and 16-bit unsigned integers (uint8 and uint16). If pixel values fall outside the range [0, 65,535], an error may be raised (strict mode) or values will be clipped (non-strict mode). Precision is automatically chosen based on image data's dynamic range.
In the future, we intend to extend support to non-Linux environments. Similarly, we intend to employ a similar approach used by openjphjs to integrate native C++ code directly into Python. While our current implementation supports the entire feature set of OpenJPH, it is not computationally optimized. We invite collaborators in the open-source community to help integrate with OpenJPH's native C++ code with direct encode/decode capabilities in Python.
If you have any questions about openjphpy, please contact us here.