Data Resources: Milestones and Building Blocks

milestone

A marker of how far one has traveled.

building block

A unit of construction or composition; something essential on which a larger entity is based.

Introduction

Data Resources—datasets and related artifacts to support scientific research and clinical implementation of artificial intelligence (AI) tools—have been an important focus of this journal since its outset. Data Resources articles describe important datasets, algorithms, standards, ontologies, policies, workflows, or data repositories that authors are making available to the community. Developing such resources is a valid scholarly effort and one worthy of recognition. Data Resources articles not only draw attention to resources of value to the radiology AI community, but also provide citable evidence of the scholarship behind such efforts and thus support further development of publicly available resources.

In their editorial, “On the Centrality of Data: Data Resources in Radiologic Artificial Intelligence,” Mongan and Halabi eloquently describe the role of public datasets and their importance to the scientific advancement in imaging AI (1). The authors are eminently qualified to discuss this topic, as their work has been central to efforts in radiology to build, evaluate, and deploy such datasets. They have planned RSNA’s numerous AI competitions and worked to promote interoperability of machine learning data as chairs of RSNA’s Artificial Intelligence committees.

With this issue of Radiology: Artificial Intelligence, I am pleased to announce that we have aggregated the journal’s Data Resources articles into a new collection, which readers can access online (2) and under the Collections tab on the journal’s web page.

Three recently published Data Resources articles highlight the value of these efforts. The Scottish Medical Imaging Archive offers more than 57 million radiology images linked to medical record information from national, population-based medical practice (3). The Duke Liver Dataset provides a valuable collection of more than 2100 abdominal MRI series from 105 patients scanned with a clinical indication of cirrhosis; there are 310 image series with manually segmented liver masks (4). The RSNA Cervical Spine Fracture CT dataset was constructed from 12 institutions across nine countries and six continents to support the 2022 Cervical Spine Fracture Detection competition, hosted by RSNA (5).

To date, the journal has published 15 diverse, robust, and valuable resources for the radiology AI community. Accompanying the collection is a “living index”: a table that lists the features of Data Resources articles published in this journal (6). Readers can sort the index by resource name, publication date, and other attributes, such as the number of patients. One can filter the resources by content codes—those two-letter abbreviations used to index articles in RSNA journals and annual meeting content, such as “NR” for neuroradiology. The journal will update this index as new Data Resource articles and their associated datasets join the collection. I commend Bardia Khosravi, MD, and Xueyan Mei, PhD, members of our Trainee Editorial Board, for their work to build and maintain this index.

We also announce a new policy related to submission of Data Resources manuscripts (7). The policy’s goal is to make related datasets and other resources publicly available to help accelerate scientific advances in our field. Going forward, authors of Data Resources manuscripts are required explicitly to place their resources in public repositories. Key aspects of the policy include:

• ■ Upon submission of a Data Resources manuscript, the authors must provide their materials in an appropriate external repository to assure that the resource is discoverable, reproducible, and reusable. A personal or laboratory website is not sufficient.

• ■ If the resource is not open to the public at the time of submission, authors must provide secure links and/or passcodes so that the editorial team and reviewers can access and evaluate the data in a confidential manner.

• ■ If the journal accepts a Data Resources manuscript for publication, the authors agree to deposit their resource (data, software, etc) in an appropriate public repository and to make the resource publicly available without restriction for at least 5 years.

To complement the Data Resources articles, we offer two Special Reports. Wilson and colleagues describe two distinct but interrelated approaches to bridge existing data silos and accelerate the pace of innovation: data liberation and crowdsourcing (8). They make a cogent argument for collaboration and sharing of data to support AI research. Bell and Shimron, in their report entitled, “Sharing Data Is Essential for the Future of AI in Medical Imaging,” vividly present the challenges and opportunities of open science (9). As they describe, the tenets of open science allow for reproducibility of research in medical imaging AI through transparent and collaborative sharing of resources such as software, data, methodology, and education. Open science presents an opportunity to stimulate innovative research and to support diverse and inclusive research environments.

Data Resources articles are milestones: They highlight the accomplishment of constructing a valuable contribution to the AI community. Those resources typically require intense collaboration across research centers, health systems, and even across continents. The projects require significant intellectual effort and deserve the “academic credit” granted by publication in a peer-reviewed journal. Data Resources articles also are building blocks: They provide the foundations to enable further exploration and innovation.

We celebrate the important contributions represented by this journal’s Data Resources articles and look forward to further growth in this critical area.