lion-ai/umie_datasets

Dataset Card for Dataset Name

UMIE (Unified Medical Imaging Ensemble) is currently the largest publicly available dataset of annotated radiological imaging, combining over 20 open-source datasets into a unified collection with standardized formatting and labeling based on the RadLex ontology.

Dataset Details

Dataset Description

UMIE datasets combine more than 20 open-source medical imaging datasets, containing over 1 million radiological images across multiple modalities (CT, MRI, and X-ray). The dataset is unique in its standardized approach to medical image data organization, using unified preprocessing pipelines and the RadLex ontology for consistent labeling across all included datasets.

This resource combines images from 12 open-source datasets, spanning X-ray, CT, and MRI modalities. The dataset includes images for both classification and segmentation tasks, with 40+ standardized labels and 15 annotation masks. We mapped all labels and masks to the RadLex ontology, ensuring consistency across datasets. UMIE datasets aim to facilitate the development of more robust and generalizable medical foundation models akin to those in general-purpose computer vision.

Due to redistribution restrictions of some opensource datasets, we release only a subset of UMIE datasets on Hugging Face. To reproduce our entire datasets, go to our repo on GitHub. In our repo, we collect the unified preprocessing pipeline that standardizes the heterogeneous source datasets into a common UMIE format, addressing challenges such as diverse file types, annotation styles, and labeling ontologies. The preprocessing scripts are modular and extensible, so that you can use existing preprocessing steps to easily incorporate new datasets.

• Curated by: TheLion.AI
  • Repository: https://github.com/TheLion-ai/UMIE_datasets
  • Paper [optional]: https://medium.com/thelion-ai/umie-datasets-83c04305b069
  • Demo [optional]: TBA

  Uses

  Direct Use

  • Training and evaluation of medical imaging AI models
  • Development of foundation models for medical imaging
  • Medical image classification and segmentation tasks
  • Research in medical computer vision
  • Benchmark dataset for medical imaging tasks

  Out-of-Scope Use

  Clinical diagnosis or medical decision-making without proper validation Applications requiring real-time processing without proper testing Use cases requiring additional modalities not included in the dataset

  Dataset Structure

  • Standardized file organization
  • Consistent image formats (converted from various sources including DICOM)
  • Unified mask formats
  • Labels following RadLex ontology
  • Unique identifiers across all datasets

  The dataset comprises of several opensource datasets. Each sub dataset is treated as a separate split. The dataset file tree looks as follows: [sub dataset ID]_[sub dataset name]->[phase name e.g."CT arterial"]->Images / Masks directory

  The information about individual imgs, such as whether it has a mask or labels is stored in a jsonl file. Each sub dataset has its own .jsonl file. You can check the json file to find which images come from the same study.

  Each image in the dataset has a unique identifier. If an image has a mask, mask has the same file name as its respective image.

  For a complete list of labels in UMIE check labels.py For a complete list of masks with their encoding check masks.py

  Dataset Creation

  Curation Rationale

  The dataset was created to address several key challenges in medical AI:

  • Lack of large-scale, standardized medical imaging datasets
  • Inconsistent formatting across existing datasets
  • Absence of common ontology for medical image annotation
  • Need for foundation models in medical imaging

  Although the number of opensource medical datasets is growing, we are lacking data formating and labeling standards. Due to the plethora of formating in the available data and lack of a common ontology for labeling, it used to be difficult to create a large-scale dataset of medical imaging. To fascilitate this process we created pipelines with reusable preprocessing steps to convert the data to a common format and a common labeling and masks ontology. This dataset collects the results of these pipelines. The pipelines are also available as opensource on our GitHub.

  Source Data

  The source data in UMIE datasets comes from opensource datasets. We provide a complete list of source datasets with links to their original source below. We did not collect any data ourselves.

  Data Collection and Processing

  The dataset combines images from 20+ open-source medical imaging datasets. Processing includes:

  • Standardized preprocessing pipelines
  • Conversion of various image formats (DICOM, PNG, etc.)
  • Mask extraction from various formats (XML, etc.)
  • Label standardization using RadLex ontology
  • Unique identifier assignment
  • Optional steps for handling missing annotations

  For preprocessing, we created custom pipelines with reusable steps, allowing to simplify the process to drag and drop. Refer to our GitHub repo for the exact code of the preprocessing pipelines.

  Who are the source data producers?

  The source data comes from various medical institutions and public medical imaging repositories, including:

  • The Cancer Imaging Archive
  • Stanford AIMI
  • The Cancer Imaging Archive
  • Grand Challenge

  Below you can find citations and links to the original sources of the datasets. We list only the datasets present on HuggingFace. Since not all source datasets in UMIE allow redistribution, some datasets requires downloading the data from source location and then use our pipelines on GitHub to preprocess it to UMIE format. 0. KITS 23

  @misc{heller2023kits21,
        title={The KiTS21 Challenge: Automatic segmentation of kidneys, renal tumors, and renal cysts in corticomedullary-phase CT}, 
        author={Nicholas Heller and Fabian Isensee and Dasha Trofimova and Resha Tejpaul and Zhongchen Zhao and Huai Chen and Lisheng Wang and Alex Golts and Daniel Khapun and Daniel Shats and Yoel Shoshan and Flora Gilboa-Solomon and Yasmeen George and Xi Yang and Jianpeng Zhang and Jing Zhang and Yong Xia and Mengran Wu and Zhiyang Liu and Ed Walczak and Sean McSweeney and Ranveer Vasdev and Chris Hornung and Rafat Solaiman and Jamee Schoephoerster and Bailey Abernathy and David Wu and Safa Abdulkadir and Ben Byun and Justice Spriggs and Griffin Struyk and Alexandra Austin and Ben Simpson and Michael Hagstrom and Sierra Virnig and John French and Nitin Venkatesh and Sarah Chan and Keenan Moore and Anna Jacobsen and Susan Austin and Mark Austin and Subodh Regmi and Nikolaos Papanikolopoulos and Christopher Weight},
        year={2023},
        eprint={2307.01984},
        archivePrefix={arXiv},
        primaryClass={cs.CV}
  }
  

  1. CoronaHack
  2. Alzheimers Dataset
  3. Brain Tumor Classification
  4. COVID-19 Detection X-Ray
  5. Finding and Measuring Lungs in CT Data
  6. Brain CT Images with Intracranial Hemorrhage Masks
  7. Liver and Liver Tumor Segmentation
  8. Brain MRI Images for Brain Tumor Detection
  9. Knee Osteoarthritis Dataset with Severity Grading
  10. Chest X-ray 14

  @inproceedings{wang2017chestx,
    title={Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases},
    author={Wang, Xiaosong and Peng, Yifan and Lu, Le and Lu, Zhiyong and Bagheri, Mohammadhadi and Summers, Ronald M},
    booktitle={Proceedings of the IEEE conference on computer vision and pattern recognition},
    pages={2097--2106},
    year={2017}
  }
  

  Due to the licencing restrictions, we were not able to publish on Hugging Face all the datasets that UMIE supports. Some datasets do not allow for redistributing tghe data in the modified format. To replicate our complete dataset, go to our GitHub Repo and use the preprocessing pipelines for the datasets listed below: 10. Brain Tumor Progression

  @article{schmainda2018data,
    title={Data from brain-tumor-progression},
    author={Schmainda, Kathleen and Prah, Melissa},
    journal={The Cancer Imaging Archive},
    volume={21},
    year={2018}
  }
  

  12. COCA- Coronary Calcium and chest CTs
  13. BrainMetShare

  Annotations [optional]

  Annotation process

  • Original annotations from source datasets are preserved
  • Labels and masks are mapped to RadLex ontology IDs
  • Consultation with radiologists for proper ontology mapping
  • Multi-label classification approach where necessary

  Who are the annotators?

  Original annotations come from the source datasets' creators. The mapping to RadLex ontology was performed by the UMIE team in consultation with radiologists.

  Personal and Sensitive Information

  The dataset follows the distribution model of ImageNet — instead of redistributing the data directly, it provides:

  • Instructions for downloading from original sources
  • Preprocessing scripts for standardization
  • Direct distribution only for datasets that allow redistribution

  Bias, Risks, and Limitations

  • Dataset quality depends on original source data quality
  • Potential biases from source dataset collections
  • Some labels may use more general RadLex IDs due to ontology limitations
  • Varying levels of annotation detail across source datasets

  Recommendations

  • Validate model performance on independent test sets before clinical use
  • Consider potential biases in source datasets
  • Review RadLex ID mappings for specific use cases
  • Check original dataset licenses for usage restrictions

  Citation [optional]

  If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->

  Dataset Card Authors

  Barbara Klaudel, Aleksander Obuchowski, Andrzej Komor, Piotr Frąckowski, Kacper Rogala, Kacper Knitter

  Dataset Card Contact

  Barbara Klaudel (team leader) LinkedIn