Session Item

Secondary dose calculation (SDC) with an independent algorithm is one option to perform plan-specific quality assurance (QA). While measurement-based QA can potentially detect errors in plan delivery, these measurements are typically only compared to calculations on homogeneous phantom geometries instead of patient CT data. We analyzed the sensitivity and specificity of a SDC software by purposely introducing different errors, providing thresholds for optimal decisions.

Thirty clinical head-and-neck VMAT plans of recent patients generated in the Philips Pinnacle treatment planning system (TPS) (version 16.2) for an Elekta Versa HD were recalculated in RadCalc (version 7.2) with its Monte Carlo algorithm. Either gamma passing rates or deviations of relevant dose parameters (target D98, D95 or D02, or clinically relevant organs at risk (OAR) parameters) were obtained. Thirty plans were purposely modified to deviate by more than 2% in at least one relevant target or OAR parameter. Errors included: overriding the density outside of an intended structure, excluding density information from the calculation by setting a wrong couch removal line, using an incorrect beam model with either the flattening filter removed or a wrong beam energy, or eight modifications with changes of output factors, penumbra or out-of-field dose. All plans were recalculated in RadCalc. Additionally, plan measurements on the ArcCHECK phantom were evaluated against TPS calculations. A receiver operating characteristics (ROC) analysis was performed to obtain optimal decision thresholds.

Area under curve (AUC) and decision thresholds from the ROC analysis are listed in Table 1. A third of the errors was neither detected by phantom QA nor by SDC. The largest AUC was obtained when basing decisions on maximum deviations of target dose parameters (max. dev. target) in RadCalc. Combining evaluation criteria from phantom measurements (gamma 2%/2mm) and SDC (max. dev. target) at thresholds of 89.4% and 2.0%, respectively, improves the sensitivity (missing only one error instead of four or nine) while keeping the same level of specificity as the individual criteria. The calculated dose to OAR often deviates considerably between RadCalc and the TPS, so the use of OAR-based criteria requires an improved beam model in RadCalc.

Table 1: AUC and decision threshold from the ROC analysis

The sensitivity and specificity of the SDC is comparable to that of phantom QA. A combination of both methods leads to higher sensitivity keeping the same specificity. Decisions based on DVH criteria should be given preference over gamma analysis. Further improvements are expected after fine-tuning the beam model in RadCalc.