AID-RT: Standardising Artificial Intelligence Documentation in RadioTherapy with a domain-specific model card

Abstract

<h2>Abstract</h2><h3>Background and Purpose</h3> Insufficient documentation of artificial intelligence (AI) models remains a widespread issue, which hampers reproducibility in research environments and safe integration in clinical departments. Our goal was to develop a standardised, structured, and domain-specific reporting framework tailored to AI models in radiotherapy (RT), enhancing transparency and accountability. <h3>Methods</h3> A working group was formed after the ESTRO Physics Workshop 2023, <i>"AI for the Fully Automated Radiotherapy Treatment Chain",</i> comprising 16 experts from 13 institutions<i>.</i> We reviewed existing initiatives for AI model and data reporting and drafted an initial template, which was sent for review to all participants. Three popular RT applications were selected to define task-specific fields: synthetic CT, segmentation, and dose prediction. Five review rounds were performed, where suggested changes were voted in a live shared Google doc. Unclear fields and conflicting votes were discussed at online meetings, and consensus was reached by majority voting. <h3>Results</h3> The final template included 6 sections: 0) Card metadata, 1) Model basic information; 2) Model technical specifications (i.e. architecture, software and hardware); 3) Training data, methodology, and information; 4) Evaluation data, methodology, and results (a.k.a commissioning for clinical models); and 5) Other considerations, including ethical use, risk analysis, and monitoring. It is publicly available at Zenodo as a Microsoft Word document and as a digital template to facilitate information entry at Streamlit.app. <h3>Conclusions</h3> We proposed a practical, consensus-driven template tailored to the unique requirements of AI models in RT, with applicability in both research and clinical environments, conveying the key information required for informed use

Ähnliche Arbeiten