A Universal Predictive DVH Modeling Toolkit
J Tan1*, L Appenzoller1, D Yang1, S Mutic1, K Moore2, (1) Washington University in St. Louis, St. Louis, MO, (2) University of California, San Diego, San Diego, CATH-A-116-4 Thursday 8:00AM - 9:55AM Room: 116
Purpose: The goal of this work was to develop an efficient DICOM-based tool to create predictive DVH (pDVH) models for treatment plan automation and knowledge-based IMRT quality control (QC) independent of treatment planning system (TPS).
Methods: The tool was implemented in MATLAB. RT dose and structure files exported from different TPSs including Eclipse, Pinnacle, and Tomotherapy were used as the input to the pDVH DICOM tool. pDVH models based on the correlation of expected dose to the distance from a voxel to the PTV surface in a cohort of patients were created for multiple treatment sites (intact prostate, prostate bed, head-and-neck, endometrial, cervical). To accelerate the computation of a distance vector field for efficient model building, multiple techniques were employed including GPUs acceleration and parallel computation using multiple CPU cores. The tool has a full GUI and can also run in batch mode with minimal user interaction. It automatically identifies outliers for treatment plan quality control and pDVH model refinement. Identified outliers can be replanned and reimported to refine the accuracy of the DVH prediction.
Results: The inclusion of GPUs reduces computation time by 200 times, requiring <1 second to compute an OAR vector field and only a few seconds to build a pDVH model. The pDVH DICOM tool has been used by multiple institutions and pDVH models for intact prostate and cervical cancer were created in less than one hour without GPU acceleration.
Conclusion: We developed a universal DICOM-based tool that provides clinicians a novel and highly efficient plan quality control mechanism by predicting OAR DVHs based on local institutional best practices. This technology can be used to asses inter-institutional and inter-TPS plan quality based on aggregate clinical data instead of contrived benchmark planning sets.
Funding Support, Disclosures, and Conflict of Interest: Supported by a research agreement with Varian Medical Systems.