KEY POINTS
- The retrospective validation included 20 consecutive head and neck cancer patients previously treated with intensity-modulated proton therapy at the Danish Centre for Particle Therapy; 18 had oropharyngeal cancer and two laryngeal cancer. All had originally qualified for protons because predicted grade 2 dysphagia and/or grade 3 xerostomia was reduced by >5 percentage points versus VMAT.
- For each patient, investigators generated two automated proton plans using a rule-based Python script in RayStation: one with a standard five-field arrangement and one reproducing the clinically selected field configuration. Clinical plans used 5–7 fields and were manually generated in Eclipse.
- Automated optimization required a median 64 minutes with either field strategy after approximately 20 minutes of initial setup. Experienced proton physicists typically required approximately 2–3 days to produce the clinical comparison plan.
- All clinical plans and all automated plans using the clinical field arrangement passed robustness criteria. By contrast, 6/20 standard five-field plans (30%) failed at least one robustness requirement; three had inadequate robust target coverage, and four exceeded the body hotspot criterion.
- Automated plans generally maintained target dose while lowering several organ doses. With clinical fields, median mean dose was reduced by 1.4 Gy to the left parotid and 1.3 Gy to the right parotid versus manual clinical plans, while glottic mean dose fell from 13.7 to 9.1 Gy and spinal-cord maximum dose from 35.9 to 16.9 Gy.
- Differences in predicted toxicity were small. Median dysphagia NTCP differed from clinical planning by only 0.1 percentage points with the standard field set and 0.3 percentage points with clinical fields; xerostomia differed by 0.3 percentage points with either approach. For most patients, automated-versus-clinical NTCP differences were within 2 percentage points.
- The important exception was geometry: four standard five-field plans had dysphagia NTCP 4.8–6.8 percentage points higher than their clinical plans. The failures occurred particularly when limited chin-to-shoulder separation meant target regions were insufficiently covered by two fields; the authors consequently modified the script to add a posterior field in such cases.
CLINICAL TAKEAWAY
For centres using model-based proton selection, automated robust IMPT planning could turn a multi-day comparative-planning task into roughly an hour while preserving the NTCP information used to select patients. A fixed beam arrangement is not completely plug-and-play: anatomical cases with poor chin–shoulder separation still require an additional posterior field, and prospective deployment is needed before assuming the workflow can replace expert review.