Adaptive cancer radiation plans generated in under 5 minutes

Summary

• The approach was tested on repeat CT scans from 10 patients: six head and neck cases and four breast cases.

• Automated daily plans were generated in about 4.5 minutes on average and differed by less than 0.5 Gy from manually created plans on most measures.

• When planning constraints conflicted, every affected case was resolved after one informed round of adjustments.

Why do radiation plans need to change during treatment?

Radiation treatment is usually delivered over multiple sessions using a plan based on the patient's anatomy at the start. During treatment, tumours can shrink and organs or internal cavities can move or change shape.

Repositioning the patient can correct an external setup difference, but it cannot fully account for internal anatomical changes. Creating a new plan each day can address those changes, but the plan must be generated within the few minutes available during a treatment session.

Why is this a prescriptive analytics problem?

A daily treatment plan requires a decision across several competing objectives: deliver enough radiation to the tumour, limit dose to surrounding organs and stay close to the clinical priorities established in the original plan.

The problem is therefore not simply to detect anatomical change. It is to determine a new set of treatment decisions under a strict time limit, and to identify which planning goals should be relaxed if the day's anatomy makes it impossible to satisfy all of them at once.

How did the partners make daily replanning faster?

Researchers in optimisation and radiation oncology from the University of Amsterdam and Massachusetts General Hospital developed a two-part strategy.

First, the method uses the original treatment plan as a guide. It looks at how much radiation should reach the tumour and how much should be kept away from nearby healthy tissue, then creates a new plan that stays as close as possible to those goals for the patient’s anatomy that day. Second, it identifies when those goals cannot all be met at the same time and shows where the conflict lies, so clinicians can make one informed adjustment instead of repeatedly changing the plan through trial and error.

The team tested the approach on repeat CT scans from 10 patients and compared the automated plans with simple patient repositioning, manually created plans from expert dosimetrists and a further-refined version of the automated plan.

What did the comparison show?

The automated plans took about 4.5 minutes to generate on average, placing them within the time window needed for online adaptive treatment.

Compared with repositioning alone, the daily plans improved tumour coverage, in some cases by several Gy. Compared with manually created plans, they differed by less than 0.5 Gy on most measures while requiring substantially less manual planning time.

The constraint-handling step resolved every case with conflicting planning goals after one round of adjustments. For patients, the research points to a route toward treatment plans that respond to daily anatomy without making the planning workload too slow for routine use.

What can the findings mean for other applications?

The next phase will involve testing the method with the lower-quality images available on an actual treatment day rather than research-quality repeat CT scans. This is necessary before conclusions can be drawn about wider clinical use.

More broadly, the approach shows how reference points and automated constraint diagnostics can support time-critical decisions with competing objectives. Similar methods could be examined in other settings where a previously agreed plan must be re-optimised quickly when operating conditions change.