2026 ESTRO Artificial Intelligence in Radiotherapy Clinical Practice: "Meet AI: Your (future?) clinical partner"

Meet AI: Your (Future?) Clinical Partner — PrécisThis lecture reframes the clinician’s relationship to AI around fiduciary duty rather than technical performance. Because physicians owe patients a duty of care, any AI that can harm patients must be monitored—but any AI that demonstrably improves care creates a potential obligation to use it. The clinician becomes an “augmented fiduciary,” not a gatekeeper or mere technical expert.The first half builds the case that humans are poor at quantifying risk. Evidence includes surgeon-versus-radiation-oncologist practice variation driven by patient sex and socioeconomic status, MDADI/DIGEST outcomes in HPV+ oropharyngeal cancer across TORS and radiation strategies, and prospective blinded studies showing expert physicians perform poorly at detecting extranodal extension on CT and osteoradionecrosis on imaging, with reported confidence uncorrelated with accuracy. By contrast, deep learning models (e.g., Kann’s ENE work, radiomics ORN pipelines) achieve supra-human discrimination.The second half addresses how AI adoption actually happens. Drawing on Choudhury’s systems-thinking framework, the BMJ Health & Care Informatics algorithm-evaluation checklist, ONC’s 2024 “FAVES” predictive-DSI guidance, and equity imperatives, the talk works through concrete examples: externally validated oncologic prediction GUIs (MDACC/UMCG/PMH, c-index ~0.71–0.76), THALIS and DITTO digital-twin visual analytics, deep Q-learning for sequential treatment selection, explainable-AI decision support, and autosegmentation—where HECKTOR benchmarks and 26-observer Dice variability (GTV 0.65–0.82) show models approaching the inter-observer ceiling. Uncertainty-aware prediction is positioned as the next step, enabling direct ISO 14971 risk estimation and a shift from QA to error-mode analysis. LLM agents for autonomous planning preview “autonomous radiotherapy.”The closing section warns of de-skilling: automation bias, eroded verification habits, degraded training ecology, and organizational amplifiers drive dependence irrespective of model accuracy. The talk ends by asking who on the team will own continuous monitoring of each AI component—advocating a process team, not merely a checklist—anchored to MSK’s AI lifecycle monitoring, ISO 14971, and FDA guidance on keeping patients safe.