AI-coded radiation therapy software requires unique quality assurance

Abstract

Software is an inseparable part of modern radiation therapy (RT), running everything from treatment planning and record-and-verify systems to the small scripts that physicists write to bridge gaps between clinical systems. Increasingly, software codes are crafted with the help of artificial intelligence (AI): large language models (LLMs) can now generate, modify, test, and document the software on which patient care depends. This prompts a question for our profession, known for its focus on quality assurance (QA) and safety: does software written with AI assistance demand a QA strategy of its own, distinct from the checks we already apply to any clinical software? This question spans the full range of software encountered in the clinic, from FDA-cleared commercial products where development QA rests with the manufacturer, to in-house scripts and tools where the physicist is often both the developer and the end user. The question is not whether such software must be verified and validated (it must), but whether AI's involvement in writing the code changes what that verification should look like, and by whom it should be performed, if it is beyond what we already require for any clinical software. To examine this emerging question, we invited two medical physicists to argue the opposing sides of the proposition: “AI-coded radiation therapy software requires unique QA.” Arguing for the proposition is Alon Witztum, DPhil, DABR; arguing against the proposition is Yi Wang, PhD, DABR.

Dr. Alon Witztum is a medical physicist and Associate Professor in the Department of Radiation Oncology at the University of California, San Francisco (UCSF). He earned his DPhil in Oncology from the University of Oxford, completed his clinical residency in medical physics at UCSF, and is board-certified by the American Board of Radiology. His work centers on data-driven innovation, clinical automation, and toxicity and outcome modeling. In his position as Director of Radiation Oncology Reporting & Analytics at UCSF he develops informatics pipelines and tools for the clinic, research, quality improvement, and operations. He has contributed to textbooks and given invited talks on the use of AI in radiotherapy QA. As part of the AAPM he serves on the Big Data Subcommittee and the working group on MPPG 20: patient-specific QA for dynamic and intensity modulated RT.

Dr. Yi Wang is a medical physicist from the Department of Radiation Oncology at Massachusetts General Hospital (MGH) and an Assistant Professor of Radiation Oncology at Harvard Medical School (HMS) in Boston, MA. He is certified in Therapeutic Radiologic Physics by the American Board of Radiology (ABR). He earned his Ph.D. in Biomedical Engineering from the University of Michigan in 2009 and completed the Harvard Medical Physics Residency Program in 2012. Since then, he has been working as a faculty physicist in the Department of Radiation Oncology at MGH. Locally, he leads an AI Lab and serves as the Director of the Mass General Brigham (MGB) Undergraduate Fellowship Program in Medical Physics. As an internationally recognized expert on AI for RT, he serves in multiple AI-related committees, task groups, and working groups in the American Association of Physicists in Medicine (AAPM), including the Machine Intelligence Subcommittee, Ad Hoc Advisory Committee on Artificial Intelligence Boot Camps (AHAIBC), the Vice Chair of Task Group 384 (Clinical Implementation of Automated Segmentation for Adaptive Radiation Therapy), the Chair of the Working Group on Generative Artificial Intelligence (WGGenAI), and the Chair of Task Group 477 (Harmonizing Terminology of Generative Artificial Intelligence in Radiation Oncology and Radiology). He authored over 40 articles and book chapters and edited two scientific books on AI. He serves as an Associate Section Editor and sits on the Editorial Board for Practical Radiation Oncology (PRO).

Dongxu Wang developed the concept and contributed to editing. Alon Witztum and Yi Wang contributed to writing and editing.

The authors declare no conflicts of interest.