Encrypted login | home

Program Information

A Statistical Voxel Based Normal Organ Dose Prediction Model for Coplanar and Non-Coplanar Prostate Radiotherapy

no image available
A Tran

A Tran*, V Yu , D Nguyen , K Woods , D Low , K Sheng , UCLA, Los Angeles, CA


SU-F-BRB-10 (Sunday, July 12, 2015) 4:00 PM - 6:00 PM Room: Ballroom B

Purpose: Knowledge learned from previous plans can be used to guide future treatment planning. Existing knowledge-based treatment planning methods study the correlation between organ geometry and dose volume histogram (DVH), which is a lossy representation of the complete dose distribution. A statistical voxel dose learning (SVDL) model was developed that includes the complete dose volume information. Its accuracy of predicting volumetric-modulated arc therapy (VMAT) and non-coplanar 4π radiotherapy was quantified. SVDL provided more isotropic dose gradients and may improve knowledge-based planning.

Methods: 12 prostate SBRT patients originally treated using two full-arc VMAT techniques were re-planned with 4π using 20 intensity-modulated non-coplanar fields to a prescription dose of 40 Gy. The bladder and rectum voxels were binned based on their distances to the PTV. The dose distribution in each bin was resampled by convolving to a Gaussian kernel, resulting in 1000 data points in each bin that predicted the statistical dose information of a voxel with unknown dose in a new patient without triaging information that may be collectively important to a particular patient. We used this method to predict the DVHs, mean and max doses in a leave-one-out cross validation (LOOCV) test and compared its performance against lossy estimators including mean, median, mode, Poisson and Rayleigh of the voxelized dose distributions.

Results: SVDL predicted the bladder and rectum doses more accurately than other estimators, giving mean percentile errors ranging from 13.35-19.46%, 4.81-19.47%, 22.49-28.69%, 23.35-30.5%, 21.05-53.93% for predicting mean, max dose, V20, V35, and V40 respectively, to OARs in both planning techniques. The prediction errors were generally lower for 4π than VMAT.

Conclusion: By employing all dose volume information in the SVDL model, the OAR doses were more accurately predicted. 4π plans are better suited for knowledge-based planning than the VMAT plans that are strongly biased in its dose gradient orientation.

Funding Support, Disclosures, and Conflict of Interest: This project is supported by Varian Medical Systems.

Contact Email: