Exhibit Hall | Forum 4
Purpose: The ability to calculate dose accurately and obtain optimal fluence efficiently is crucial to apply machine learning to treatment planning. To manage the computational load the dose calculation kernel is often truncated which may influence the resulting optimized fluence. The aim of this study is to determine such influence to guide the parameter setting for use in the machine learning algorithms.
Methods: Using a pencil beam dose calculation kernel, the error introduced in truncation is propagated into the fluence optimization problem. A metric is derived giving the upper bound on the error for a particular beamlet for some truncation radius. The sensitivity matrix is then defined as the maximum error introduced for a beamlet given a change in truncation radius. Tests were performed on a set of 50 Pancreas and 15 Head and Neck treatment sites for varying truncation radii and beam energies. The resulting differences in the calculated optimal dose and fluence were measured.
Results: The sensitivity of the optimization problem increases with decreasing truncation radius. The cross-correlation remained significantly high (>0.99) while the mutual information decreased with truncation radius (0.7 to 1.0). The beam energy did not have any significant effect on the dose distributions or resulting fluence maps. The resulting optimal dose distributions had very little differences between truncation radii with gamma passing rates around 99% for 1%/1mm criteria. Due to the size of the target and required beamlet resolution, the head and neck cases could only be performed with radii less than 10 mm but with significant error at 5 mm.
Conclusion: The errors introduced through kernel truncation in the dose calculation propagate into the fluence optimization problem and can be quantified. The magnitude of the errors vary between treatment site, have no apparent energy dependency, and can be bounded above using the sensitivity matrix.
Not Applicable / None Entered.
TH- External Beam- Photons: IMRT/VMAT dose optimization algorithms