(Wednesday, 8/1/2018) 10:15 AM - 12:15 PM
Room: Davidson Ballroom A
Purpose: Many knowledge-based planning and automated planning techniques rely on accurate dose prediction, which is critical to ensure optimal dose outcomes. We directly compare the dose prediction accuracy of statistical voxel dose learning (SVDL), spectral regression (SR), and support vector regression (SVR).
Methods: SVDL, SR, and SVR were used to predict the dose of 4Ï€ and VMAT head and neck, 4Ï€ lung, and VMAT prostate plans. 20 cases of each site were used for k-fold cross-validation, with k=4. SVDL bins voxels by their Euclidean distance to the PTV and takes the median to predict the dose of new voxels. PTV distance, various combinations of the distance components, PTV size, OAR size, and prescription dose were used as features for SR and SVR. 28 features in total were included. Principal component analysis (PCA) was performed on the input features to test the effect of dimension reduction. For the coplanar VMAT plans, separate models were trained for voxels within the same axial slice as PTV voxels and voxels outside of the primary beam. The effect of training separate models for each OAR compared to all voxels collectively was also tested. The mean squared error was calculated to evaluate the voxel dose prediction accuracy.
Results: SVDL using separate models for each OAR had the lowest MSE for all sites and modalities: 15.3 (HN 4Ï€), 10.3 (HN VMAT), 6.20 (Lung 4Ï€), and 5.51 (Prostate VMAT). For the machine learning methods, PCA consistently reduced the 4Ï€ prediction error and was most effective using all/most of the possible principal components. Separate OAR models were more accurate than training on all OAR voxels in all cases.
Conclusion: Despite the use of more sophisticated machine learning methods in combination with dimension reduction, SVDL is more robust to patient variability and provides the most accurate dose prediction method.
Funding Support, Disclosures, and Conflict of Interest: NIH U19AI067769 DE-SC0017687 NIH R21CA228160 DE-SC0017057 NIH R44CA183390 NIH R43CA183390 NIH R01CA188300
Keywords
Taxonomy
TH- Dataset analysis/biomathematics: Machine learning techniques
Contact Email
