AI in Radiation Therapy Planning

Radiation therapy tries to deliver a high dose to cancer while sparing healthy tissue. Planning starts with contouring: outlining the visible tumor (gross tumor volume, GTV), areas of likely microscopic spread (clinical target volume, CTV), a margin for setup and motion (planning target volume, PTV), and organs at risk (OARs) such as the heart, lungs, spinal cord or rectum. Manual contouring can take hours for complex sites like head and neck, and different clinicians draw noticeably different outlines. Auto-contouring was first done with atlas-based methods, which warp outlines from reference patients onto a new scan. Deep learning models, often U-Net-style networks, now generally perform better and are built into commercial systems from vendors such as Varian, Elekta, RaySearch, Siemens Healthineers and Mirada. Microsoft's InnerEye research project, developed with Addenbrooke's Hospital in Cambridge, released open-source tools in this area. OAR contouring is the most mature use. Target contouring is harder because it depends on clinical judgment, pathology and imaging beyond the scan. Dose planning uses AI in two main ways. Knowledge-based planning, such as Varian's RapidPlan, learns from prior approved plans to predict achievable dose-volume histograms for a new patient. Deep learning dose prediction and automated planning go further by estimating the 3D dose distribution and driving the optimizer. Adaptive radiotherapy depends on this speed. Systems such as Varian Ethos and MR-linacs like Elekta Unity re-image the patient at treatment and adapt the plan. A common misconception is that AI plans treatment on its own. Physicians approve contours and prescriptions, and medical physicists check plans and run quality assurance before delivery.