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

## TL;DR

This paper introduces a standardized model card for documenting AI in radiotherapy, developed by experts to improve transparency and safe clinical use.

## Contribution

A consensus-based, domain-specific model card template for AI in radiotherapy, tailored to enhance reproducibility and clinical integration.

## Key findings

- The template includes six sections covering metadata, model specs, training, evaluation, and ethical considerations.
- The template was developed through five review rounds and majority voting among 16 experts from 13 institutions.
- The model card is publicly available as a downloadable document and interactive web form.

## Abstract

•Consensus-based model card for documenting artificial intelligence in radiotherapy.•Initiative from ESTRO Physics Workshop 2023.•Covers model and data aspects for training, evaluation, and ethical use.•Enhances transparency and safe clinical integration of artificial intelligence.•Publicly available template for research, clinical, and educational use.

Consensus-based model card for documenting artificial intelligence in radiotherapy.

Initiative from ESTRO Physics Workshop 2023.

Covers model and data aspects for training, evaluation, and ethical use.

Enhances transparency and safe clinical integration of artificial intelligence.

Publicly available template for research, clinical, and educational use.

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.

A working group was formed after the ESTRO Physics Workshop 2023, “AI for the Fully Automated Radiotherapy Treatment Chain”, comprising 16 experts from 13 institutions. 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 shared online document. Unclear fields and conflicting votes were discussed at online meetings, and consensus was reached by majority voting.

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 as a downloadable document template and as an interactive web-based form to facilitate information entry.

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.