Room: ePoster Forums
Purpose: To show the influence of utilizing large training datasets, mined with the Eclipse Scripting API, in knowledge-based treatment planning with RapidPlanâ„¢.
Methods: A novel software solution was developed to use the Eclipse Scripting API (ESAPI) to locate and extract plan data for multiple, large RapidPlan training sets. Filters were developed to look for plans based on treatment date, fractionation, structures, and approval status. All treatment plans that were Treatment Approved, of similar fractionation, and contained the minimum amount of structures were chosen. A structure matching algorithm was used to match structure names with the RapidPlan model structure names. The plans were exported via DICOM, anonymized, then reimported. A subset of plans was reserved for validation of the resulting models. These plans were copied, the DVH Estimation for the model was applied, calculated, and normalized to the original plan prescription. The original, clinical dose constraints were used to evaluate the resulting plan quality of the test cases.
Results: Overall, the plan quality of the resulting RapidPlan produced treatment plans were comparable or better in meeting the dose constraints compared the original plan. Above a certain threshold, more treatment plans didnâ€™t result in a marked increase in meeting dose objectives; however, without enough plans in a training set, not all organs at risk may be properly trained. The other benefit of the large training set was that less focus had to be taken to remove outliers to obtain acceptable model statistics.
Conclusion: Without utilizing tools for analyzing, locating, and extracted data for RapidPlan training sets, it would be impractical for clinical sites to build their own large, custom models. Training sets that exceed the recommended number of plans by the vendor help the model building and validation process by reducing the effort in localizing outliers.