Encrypted login | home

Program Information

Machine Learning Using Rich Geometrical Features for Coplanar VMAT Dose Prediction

no image available
A Landers

A Landers*, D Ruan , F Scalzo , K Woods , K Sheng , UCLA School of Medicine, Los Angeles, CA


TH-CD-205-6 (Thursday, August 3, 2017) 10:00 AM - 12:00 PM Room: 205

Purpose: It has been shown in knowledge-based planning methods for coplanar plans that dividing in-beam and out-of-beam voxels into separate training sets, in addition to the use of voxel Euclidian distances, can improve dose prediction accuracy due to the anisotropic dose distribution. Machine learning regressions can further expand knowledge-based dose prediction by including additional geometrical parameters. This study compares two supervised machine learning regressions, spectral regression (SR) and support vector regression (SVR) and assesses the implication of in/out-of-beam splitting for predicting coplanar VMAT dose.

Methods: 30 clinical coplanar liver VMAT plans were divided into equally-sized training and test sets. The features for each voxel included various components of Euclidean distance from the PTV, liver and PTV size, and prescription dose. The first experiment compared SR and SVR with all voxels as one training set. In this case, polar angles from the PTV centroid and from the linac source were included as features. The second experiment compared SR and SVR with voxels separated into in-beam and out-of-beam training sets. Tikhonov regularization and a radial-basis-function kernel (LIBSVM) were used for the SR and SVR, respectively. Voxel dose prediction accuracy was evaluated by the root-mean-squared-error (RMSE).

Results: When using one training set, RMSE was 0.189 and 0.165 for SR and SVR, respectively. SVR using two training sets separated by in/out-of-beam achieved the lowest RMSE of 0.132 compared to 0.145 with SR. SR took ~1 second for both training and prediction, whereas SVR took ~1 minute for training and ~10 minutes for prediction of the 15 patient testing set.

Conclusion: In addition to separating voxels into in-beam and out-of-beam training sets, machine learning can take advantage of more geometrical features. Between the two machine learning methods, SVR is more robust and accurate compared to SR. However, SR is advantageous for time-sensitive applications.

Funding Support, Disclosures, and Conflict of Interest: DOE DE-SC0017057 NIH R44CA183390 NIH R01CA188300 NIH R43CA183390 NIH U19AI067769

Contact Email: