A subcommittee of the RAPNO (Response Assessment in Pediatric Neuro-Oncology) consortium that is focused on artificial intelligence (AI-RAPNO) has released guidance on the responsible implementation of AI in pediatric neuro-oncology in the form of a two-part policy review published in The Lancet Oncology.
“These recommendations offer a practical roadmap to move from promising research to safe, equitable bedside use,” said Anahita Fathi Kazerooni, PhD, lead author of the policy article and senior author of the review, co-lead of AI-RAPNO, and Assistant Professor in the Center for Data Driven Discovery in Biomedicine (D3b) within the Division of Neurosurgery at the Children's Hospital of Philadelphia (CHOP), and the Department of Neurosurgery at the University of Pennsylvania. “As AI evolves, standardization, validation and transparency will be key to realizing personalized care for children with brain tumors.”
Need for the Guidelines
RAPNO criteria for evaluating treatment response address the distinctive challenges of pediatric central nervous system (CNS) tumors, which differ fundamentally from adult CNS tumors.
Progress has been made in implementing AI approaches and tools for imaging and radiology in adult oncology, but has yet to be applied to pediatric oncology. AI-RANO recommendations were created for adult neuro-oncology to help determine how AI can be used in clinical decision- making, and such a framework is also necessary in pediatric neuro-oncology.
International experts from the AI-RAPNO consortium subcommittee reviewed existing AI applications as they applied to pediatric neuro-oncology and created a set of recommendations for how AI can be used to aid in tasks and workflows to improve imaging analysis and treatment response assessment. By implementing these guidelines, it would enable safer treatments for young patients and more accurate and reproducible results for comparisons across trials.
The subcommittee experts conducted a detailed search of PubMed, MEDLINE, and Google Scholar for articles published between January 2000 and December 2024 that related to pediatric brain tumors, AI, machine learning, radiomics, and more. These articles included peer-reviewed original research, meta-analyses, and systemic reviews for AI or machine learning approaches in pediatric CNS tumors. A total of 125 articles were identified and reviewed.
Part one of the policy review assessed promising use cases and tools for AI. The second part focused on the benefits, challenges, and future directions of AI use in clinical practice as these tools are translated into standard use.
Challenges With AI Implementation
The authors of the policy review noted several challenges inherent in the application of AI models for pediatric neuro-oncology, including variability in imaging protocols leading to inconsistent interpretation, few annotated datasets for training reliable AI models, and regulatory and ethical considerations of integration.
Members of AI-RAPNO stressed that collaboration is needed among clinicians, researchers, and regulatory bodies to overcome such challenges.
Key Recommendations
The panel of experts recommended that imaging protocols should be standardized across institutions to reduce the variability between methodologies, and robust frameworks should be made for validating models for use in clinical settings. Infrastructure should also be adapted to prepare for AI integration, including that of electronic health records and clinical trials.
For imaging standardization specifically, the AI-RAPNO group recommends consistently defining and detailing tumor subregions and outlining areas on image results for greater accuracy and less interobserver variability.
When measuring responses to treatment, volumetric analysis is needed for irregular tumors but can be optional for other tumors. Additionally, assessment should include clinical, laboratory, and imaging assessment data to provide a truly comprehensive picture for better decision-making.
The panel also recommended that transfer learning and self-supervised learning can assist in training models when there is a scarcity of pediatric CNS tumor data, as well as the use of synthetic control groups.
AI tools created specifically for use in pediatric populations should be validated across different groups of children with consideration of various ages, tumor types, tumor locations, and imaging methodologies. Most tools created for use in adult neuro-oncology is therefore not suitable for pediatric applications. Additionally, image intensities should be aligned to ensure the extraction of reproducible radiomic features.
The AI-RAPNO experts also stressed that the AI tools should be generalizable, transparent, and equitable with vital safeguards and guardrails put into place. With all of these recommendations, AI could enhance diagnostic assessments, prediction of recurrence and progression, guide clinical decision-making, evaluate treatment-related adverse events, inform survivorship needs, and more for pediatric patients with CNS tumors.
“These guidelines help clinicians understand when and how to trust AI outputs, what evidence to look for, and how to integrate tools with RAPNO-aligned decision-making,” said Ali Nabavizadeh, MD, senior author of the policy article and co-senior author of the review, co-lead of AI-RAPNO, Director of the Translational Imaging Research Unit at CHOP’s D3b and Associate Professor at the University of Pennsylvania.
Disclosure: For full disclosures of the study authors, visit thelancet.com.
