Purpose: The utility of complexity metrics has been assessed for IMRT and VMAT treatment plans, but this analysis has never been performed for CyberKnife (CK) plans. The purpose of this study is to perform a complexity analysis of CK MLC plans, adapting and computing complexity indices previously defined for IMRT plans. Metrics were used to compare the complexity of plans created by two optimization systems and to study correlations between plan complexity and patient-specific quality assurance (PSQA) results. Relationships between pairs of metrics were also analyzed to get insight into possible interdependencies.
Methods: Two independent in-house software platforms were developed to compute six complexity metrics: modulation complexity score (MCS), edge metric (EM), plan irregularity (PI), plan modulation (PM), leaf gap (LG), and small aperture score (SAS10). MCS and PM definitions were adapted to account for CK plans characteristics. The computed metrics were used to compare the existing optimization algorithms (sequential and VOLO) in terms of plan complexity over 24 selected cases. Metrics were then computed over a large number (103) of VOLO SBRT clinical plans from different treatment sites, mainly liver, prostate, pancreas, and spine. Pearson's r was used to study relationships between each pair of metrics. Correlation between complexity indices and PSQA results expressed as gamma index passing rates (GPR) at (3%, 1 mm) and (2%, 1 mm) was finally analyzed. Correlation was regarded as weak for absolute Pearson's r values in the range 0.2-0.39, moderate 0.4-0.59, strong 0.6-0.79, and very strong 0.8-1.
Results: When compared to VOLO, sequential plans exhibited a higher complexity degree, showing lower MCS and LG values and higher EM, PM and PI values. Differences were significant for 5/6 metrics (Wilcoxon P < 0.05). The analysis of VOLO clinical plans highlighted different degrees of complexity among plans from different treatment sites, increasing from liver to prostate, pancreas, and finally, spine. Analysis of dependencies between pairs of metrics showed a very strong significant negative correlation (P < 0.01), respectively, between MCS and PM (r = -0.97), and EM and LG (-0.82). Most of the remaining pairs showed moderate to strong correlations with the exception of PI, which showed weaker correlations with the other metrics. A moderate significant correlation was observed with GPR values both at (3%, 1 mm) and (2%, 1 mm) for all metrics except PI, which showed no correlation.
Conclusions: Modulation complexity metrics were computed for CK MLC-based plans for the first time and some metrics' definitions were adapted to CK plans peculiarities. The computed metrics proved a useful tool for comparing optimization algorithms and for characterizing CK clinical plans. Strong and very strong correlations were found between some pairs of metrics. Some significant correlations were found with PSQA GPR, indicating that some indices are promising for rationalizing and reducing PSQA workload. Our results set the basis for evaluating new optimization algorithms and TPS versions in the future, as well as for comparing the complexity of CK MLC-based plans in multicenter and multiplatform comparisons.
Keywords: CyberKnife; MLC; complexity metrics; plan complexity; quality assurance.
© 2020 American Association of Physicists in Medicine.