Speakers
Description
Accelerators used in cancer radiotherapy share the physical principles of research machines but operate under constraints that fundamentally reshape how artificial intelligence can be deployed: stringent uptime requirements driven by patient scheduling, layered hardware safety interlocks that define the boundary of AI intervention, clinical regulatory obligations (IEC 62304, EU AI Act), and fault events rare by design — inverting the data-abundance assumption on which standard machine learning relies.
This lecture covers AI applications on the accelerator side — the machine that produces and conditions the therapeutic beam — and is explicitly scoped to exclude treatment planning and patient-facing clinical AI. Three application categories are treated: unsupervised fault detection on RF cavity waveforms via convolutional autoencoders; automated beam tuning via Bayesian optimisation on a differentiable beam-physics surrogate; and neural-network surrogate models extended to real-time digital twins. Open challenges — explainability, cross-facility transfer, and control-system integration — are framed as open accelerator-physics problems accessible to early-career practitioners.