Source-linked AI summary
OPTIMAM Mammography Image Database: a large scale resource of mammography images and clinical data
Mark D Halling-Brown, Lucy M Warren, Dominic Ward, Emma Lewis, Alistair Mackenzie, Matthew G Wallis, Louise Wilkinson, Rosalind M Given-Wilson, Rita McAvinchey, Kenneth C Young
TL;DR
Medical imaging AI research is constrained by the limited availability of large, shareable, well-annotated image databases. This paper describes OMI-DB, a longitudinal mammography resource integrating images with clinical, pathological, and lesion annotations. The resulting database is a valuable sharable resource containing processed and unprocessed images, annotated cancers, and clinical details.
Problem
Medical imaging research and AI development face a lack of large medical-image databases that can be shared with other researchers.
Method
OMI-DB integrates mammography images with clinical, pathological, and expert lesion annotations across relational databases and cloud storage.
Results
OMI-DB holds processed and unprocessed mammography images with annotated cancers and clinical details.
Takeaways & Limitations
Previous screening events and interval cancers support research evaluating whether abnormalities could have been detected on earlier images.
Takeaways & Limitations
Creating and maintaining an annotated mammographic database with sharing protocols is time consuming and challenging.
Abstract
from arXiv · showhide
A major barrier to medical imaging research and in particular the development of artificial intelligence (AI) is a lack of large databases of medical images which share images with other researchers. Without such databases it is not possible to train generalisable AI algorithms, and large amounts of time and funding is spent collecting smaller datasets at individual research centres. The OPTIMAM image database (OMI-DB) has been developed to overcome these barriers. OMI-DB consists of several relational databases and cloud storage systems, containing mammography images and associated clinical and pathological information. The database contains over 2.5 million images from 173,319 women collected from three UK breast screening centres. This includes 154,832 women with normal breasts, 6909 women with benign findings, 9690 women with screen-detected cancers and 1888 women with interval cancers. Collection is on-going and all women are followed-up and their clinical status updated according to subsequent screening episodes. The availability of prior screening mammograms and interval cancers is a vital resource for AI development. Data from OMI-DB has been shared with over 30 research groups and companies, since 2014. This progressive approach has been possible through sharing agreements between the funder and approved academic and commercial research groups. A research dataset such as the OMI-DB provides a powerful resource for research.
Introduction
AI development for breast screening depends on well-curated image databases. OMI-DB was created as a centralized, fully annotated dataset for mammography research.
- AI software development for breast screening relies on well-curated image databases.
- OMI-DB was created to provide a centralized, fully annotated dataset for research.
- The database initially supported Cancer Research UK projects evaluating factors affecting breast cancer detection in mammograms.
Image database: Image collection and design
OMI-DB combines automated and stand-alone collection processes with relational databases and cloud storage. It links searchable mammography metadata to clinical, pathological, and expert lesion annotations.
- Image collection: OMI-DB uses automated remote-site and stand-alone processes to collect images and clinical data identified through NBSS queries.
- Image collection: Imaging and screening data are pseudonymised, uploaded to cloud storage, and logged at collection sites.
- Data model: The database integrates radiological, clinical, and pathological information across relational databases and cloud storage systems.
- Annotation: DICOM tags create a searchable index, while experienced readers marked 7143 lesions and recorded appearance and conspicuity attributes.
Content of OMI-DB
OMI-DB contains longitudinal mammography data from three UK screening sites, including processed and unprocessed images, cancer annotations, and clinical information. Prior screening episodes and interval cancers support research on earlier detectability.
- Database content: Images and clinical data were collected from three UK screening sites, with 2,889,312 total images in the database.
- Database content: The database includes both unprocessed and processed images for studying image processing, imaging-system parameters, clinical performance, and CAD.
- Cancer data: Table 1 reports invasive status and grade distributions for screen-detected and interval cancers using NBSS clinical annotations.
- Longitudinal data: Previous screening events and interval cancers enable evaluation of whether abnormalities could have been detected earlier through altered processing or perception.
- Modalities: Collection expanded from 2D digital mammography to additional modalities including tomosynthesis and MRI.
Data Sharing
OMI-DB sharing is governed by ethical approval, funder-controlled agreements, and additional de-identification for external recipients. Secure transfer and metadata tools support sharing with academic and commercial groups.
- Governance: The database has ethical approval, and Cancer Research UK controls sharing agreements with approved academic and commercial research groups.
- Governance: External sharing requires further de-identification and dedicated records of shared cases, investigators, and access information.
- Secure access: A download-coordinating tool securely transfers data and synchronizes metadata defined by access lists.
- Research access: An open-source Python package provides an API and tools for metadata extraction and filtering of shared OMI-DB data.
Use of database
OMI-DB has supported research on breast cancer detection, imaging factors, and AI algorithms, including both algorithm training and independent evaluation.
- OMI-DB data supported virtual clinical trials examining detector type, dose, and image processing effects on breast cancer detection.
- Over 30 academic, research, and commercial groups received selected OMI-DB data to train AI algorithms.
- Independent OMI-DB images were shared to evaluate AI algorithms from prototypes through CE marked products.
Discussion
OMI-DB is a large, shareable mammography resource whose ongoing updates and longitudinal cases support research on cancer development and earlier detection.
- Creating an annotated mammographic database with sharing protocols is difficult, costly, time consuming, and challenging.
- Ongoing collection updates each case with new information and subsequent screening episodes, under sharing protocols for researchers worldwide.
- Previous screening events and interval cancers enable research into whether abnormalities were visible on earlier images.
- Sequential normal cases can be analyzed with quantitative imaging features when later malignancy is known.
- The database contains processed and unprocessed mammography images with annotated cancers and clinical details.
Funding Information
The OPTIMAM image database was funded by Cancer Research UK, with participating screening centres and contributors acknowledged for providing images and lesion annotations.
- Cancer Research UK funded the creation and development of the OPTIMAM image database.
- Three breast screening centres contributed images, while named clinical contributors annotated lesions within OMI-DB.