Multiple treatment plans could make proton therapy for cancer more reliable

Summary

• The method was tested on CT data from five head and neck cancer patients, with 57 uncertainty scenarios per patient.

• Compared with conventional single-plan robust optimisation, the approach increased the worst-case minimum tumour dose by an average of 4.52 Gy.

• Around 15 pre-planned options - about a quarter of the 57 scenarios - captured nearly all of the achievable benefit.

  
  

Why can one robust treatment plan be too cautious?

Proton therapy can concentrate radiation in a tumour while limiting exposure to surrounding tissue, but small positioning or beam-range errors can change where the dose is delivered. Treatment planners therefore account for many possible error scenarios before treatment begins.

A common approach creates one plan that must perform acceptably across all simulated scenarios. This can make the plan conservative: it protects against unlikely combinations of errors but may reduce the dose that can reliably be delivered to the tumour.

Why is this a prescriptive analytics problem?

The decision is not only how to predict uncertainty, but what treatment plan should be used when that uncertainty becomes clearer. K-adaptability addresses this by preparing a limited set of plans in advance and choosing among them later, once the patient's treatment-day position is known.

The challenge is computational. Finding the best possible set of plans is mathematically hard, so the research focuses on how to make a strong decision under uncertainty without requiring an exact solution that is too slow to calculate.

How did the team develop a practical multi-plan approach?

Optimisation researchers from the University of Amsterdam, the University of Hong Kong, and radiation oncology specialists from Massachusetts General Hospital developed a heuristic that generates a strong set of candidate plans without solving the full problem exactly.

Instead of grouping uncertainty scenarios only by physical similarity, the method groups them according to how well a treatment plan performs across them. It then builds and revisits a pool of candidate plans to identify a combination that performs well across the scenarios.

The team tested the approach on five head and neck cancer patients treated at Massachusetts General Hospital. For each patient, 57 scenarios represented realistic combinations of setup and proton-range uncertainty. The results were compared with conventional single-plan optimisation and three alternative clustering methods.

What did the comparison show?

The K-adaptability approach increased the worst-case minimum tumour dose by an average of 4.52 Gy compared with conventional single-plan robust optimisation. In practical terms, the result reduces the risk that uncertainty leaves part of the tumour receiving less radiation than intended under adverse conditions.

The team also found that about 15 plans were typically enough to capture nearly all of the benefit available across the 57 scenarios. This matters for clinical feasibility: preparing and reviewing several plans creates additional work, so the number of plans needs to remain manageable.

The method also produced higher treatment quality than the three comparison approaches used in the study, while requiring fewer planning computations than some alternatives.

What can the findings mean for other applications?

A multi-plan approach could support treatment models in which several plans are quality-checked before treatment and the most appropriate one is selected each day. Before clinical adoption, further research is needed on how to automate quality assurance and how the method performs with real daily patient images rather than simulated uncertainty scenarios.

The underlying decision structure also appears in other settings where a small set of options can be prepared before uncertainty is resolved and one option selected later. Therefore, it could also be applied to fields such as logistics and resource planning as potential areas for further application.