Session Item

The implementation of RapidPlan™, a knowledge based planning solution by Varian Medical Systems, for head and neck (H&N) radiotherapy treatment plan optimisation has been shown to have the potential to improve treatment plan quality while improving planning efficiency. The impact of the number of plans used to train a H&N RapidPlan™ model on plan quality is a parameter which has not been widely investigated. This has been assessed during an investigation along with potential strategies for implementation.

Four H&N RapidPlan™ models were trained using varying numbers of plans(150, 120, 80 and 40) with primary dose level between 60Gy and 70Gy. The plan quality achieved by the models was assessed using 29 anonymised, previously treated test patients. Quantitative assessment was carried out using clinical planning metrics and a quality metric developed during the investigation based on clinical planning aims. Significance testing was carried out on the median differences for each considered metric between matched observations using the Wilcoxon signed-rank test.

There was no statistically significant difference between the overall quality of plans produced by each model and the majority of individual planning metrics. Comparing each model to the corresponding clinical plans demonstrated a statistically significant improvement in plan quality for each model (p<0.05). The model trained with the least number of plans (40) resulted in median increases in Brainstem D_0.1cc and left parotid D_mean relative to clinical plans when compared to the other models (>2.2Gy and >0.7Gy respectively), though these results were not statistically significant.

The use of RapidPlan™ generated objectives as a starting point for manual optimisation demonstrated an overall improvement in plan quality (p<0.001) relative to clinical plans. Compared to clinical plans, this method exhibited a statistically significant reduction in median D_0.1cc to the canal(p<0.001) of 3.4Gy and brainstem(p<0.001) of 8.5Gy while improving target coverage(p<0.05). 

There was insufficient evidence to suggest that the number of plans used to train a RapidPlan™ model impacts overall plan quality. For many OAR planning metrics, a small median increase was generally observed across the testing set when using the model trained with the least plans. Despite this, these results demonstrate that a large atlas of plans is not necessarily required to produce a H&N RapidPlan™ model.

The implementation of a RapidPlan™ model to replace the Eclipse™ objective template as a starting point for planners has the potential to improve H&N plan quality. This study has demonstrated this approach maintains or improves OAR sparing relative to clinical plans while improving target coverage. In addition to plan quality, planning efficiency is also an important consideration and the investigation into this is ongoing.