Towards a Formal Description of Artificial Intelligence Models and Datasets in Radiology

The Radiology Model and Dataset Ontology (RMDO) has been created to provide formal descriptions for the growing number of artificial intelligence (AI) models and datasets in diagnostic radiology. RMDO builds upon generalized “model cards” and “datasheets for datasets” by highlighting features specific to radiology and by referencing concepts from related external ontologies and coding schemes, such as RadLex, the LOINC/RSNA Radiology Playbook, and radiology common data elements. RMDO also incorporates the tasks and methods hierarchies from the Papers With Code web resource. In accordance with the FAIR guiding principles, application of the ontology will allow AI resources to be more readily discoverable and reusable. Its application also is expected to improve the ability to match AI models with relevant datasets and to facilitate detection of potential biases in released AI models.