ROADMAP Sets Metadata Standard for Medical AI

Artificial intelligence tools in medicine increasingly require clear metadata that can support consistency, discovery, interoperability and clinical deployment. Current documentation approaches, including model cards and datasheets for datasets, capture important details but do not provide a universally accepted standard that covers both AI models and datasets. A 2026 publication in Radiology: Artificial Intelligence introduces the Radiology Ontology of AI Datasets, Models and Projects, known as ROADMAP. The ontology formally defines attributes of medical AI resources and the allowable values used to describe them. It combines human-readable terms with machine-computable structure, supports multimodal resources including medical images, structured data and unstructured text, and connects AI documentation with established biomedical terminology for more consistent description across systems, institutions, studies and clinical workflows.

A Shared Vocabulary for AI Resources

ROADMAP provides a unified set of terms for describing AI models, datasets and projects. Its structure includes synonyms, abbreviations, explicit definitions and supporting references, while its hierarchy connects broader and more specific concepts. The ontology expresses its axioms in Web Ontology Language, a standardised logic-based language that enables computational interpretation. Its design supports automated exchange and interpretation of information about AI resources.

Must Read: Unlocking Health Data Demands Governance Beyond Compliance

The ontology specifies 514 classes, including 343 with definitions, as well as 20 object properties, one data property, 717 class–subclass relationships and 773 logical axioms. The RSNA Radiology AI Data Standards Committee maintains the ontology. At the top level, ROADMAP entity includes model, dataset and project. A project can aggregate one or more models or datasets, while descriptor fields capture general information such as name, version, licence, repository link, contact information, dates, funding, ethical review, organisation, contributors and references.

ROADMAP also supports indexing through free-text keywords and established terminology systems. It links AI resources with RadLex, SNOMED CT and content codes based on RSNA specialty categories. These connections allow resources to carry terms for anatomy, diseases, medical interventions, imaging modalities, image acquisition characteristics, procedures and clinical conditions.

Model and Dataset Descriptors

The model class supports structured description of machine learning models. Model property fields capture the clinical or technical problem that a model addresses, evaluation metrics, code links and operational details needed to run the model. Additional fields describe model architecture, intended uses, ethical considerations, inputs and outputs. Model input and output descriptions can reference radiology common data elements defined in the RadElement data dictionary maintained by the RSNA and the American College of Radiology.

ROADMAP separates intended, out-of-scope and excluded model uses. It also separates intended, out-of-scope and excluded users, with user types including radiologist and referring physician. The ontology incorporates 18 performance criteria and a set of 207 graphical, matrix and scalar metrics, including precision–recall curve, confusion matrix and positive predictive value.

The dataset class describes data used to train or test an AI model. Dataset subtypes distinguish training, optimisation and testing datasets, with testing datasets further separated into internal and external forms. Dataset properties include confidentiality, external data, missing information, motivation, noise, partitioning scheme, reidentification, relationships between instances, sampling and sensitive data. ROADMAP can describe subsets by age, sex, race or ethnicity, geography, social determinants of health and clinical conditions. It can also record counts for patients, examinations, images, imaging series and physical imaging sites. Image properties cover modality, procedure, image characteristics and file format, including DICOM, JPEG, MPEG, NiFTI and PNG.

Applications, Limits and Future Development

ROADMAP terminology underpins the RSNA Annotated Library of AI Systems, known as ATLAS. ATLAS stores index cards encoded in JavaScript Object Notation, with structured information about AI models and datasets. As of December 2025, ATLAS included cards for 70 models and 77 datasets. The website provides interactive tools to create, edit and validate schema-conformant model and dataset cards, while an automated system generates preliminary cards by extracting information from published content.

ROADMAP has generated metadata for models and datasets from medical imaging AI publications, RSNA AI challenges and Medical Open Network for AI models. The ontology captures dataset composition, annotation methodology, model architecture, intended uses and ethical considerations. This structure supports Findable, Accessible, Interoperable, Reusable principles by enabling automated search, validation and reuse of metadata about AI resources and the underlying resources themselves.

The ontology has limits. Standardised descriptors may need revision as AI advances. Regulatory expectations may add requirements for lifecycle management, drift monitoring, change control or real-world performance. ROADMAP includes descriptors relevant to fairness and potential bias but does not enforce best practices for bias mitigation or algorithmic fairness. Future development focuses on expanded software tools, alignment with transparent AI reporting guidelines, support for foundation models and agent-based systems, collaboration with other specialties and integration with regulatory science tools and clinical systems.

ROADMAP gives medical AI a structured framework for describing models, datasets and projects. Its unified terminology connects radiology-specific concepts with broader AI documentation structures, while ATLAS demonstrates how the ontology can support documentation of AI resources in practice. The framework also clarifies where metadata can capture intended uses, dataset characteristics, performance metrics, ethical considerations and potential bias. Continued development keeps the framework aligned with advances in AI methods, regulatory needs and clinical applications.

Source: Radiology: Artificial Intelligence

Image Credit: iStock