(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: Small field dosimetry is significantly different from a broad beam due to loss of electron side scatter equilibrium, source occlusion, and effects related to the choice of detector. However, with the increasing use of small fields due to growing indications for IMRT and SBRT, the need for accurate dosimetry is ever more important. Here we propose to leverage machine learning strategies to reduce the uncertainties and increase the precision in determining small field output factors.
Methods: Linac output factors from a Varian TrueBeam were collected at various multi-leaf collimator (MLC) positions, jaw positions, and with and without contribution from the leaf-end transmission. The fields were defined by the MLC with the jaws and various positions. The machine learning problem was formulated as an output of random forest regression. The acquired data sets were split into ¾ for training and ¼ for testing. Absolute percentage relative error (pRE) was used to compare predicted values with ground truth. Small field output factors at various settings were also predicted using linear models trained with and without regularization as a means for comparison.
Results: Accurate predictions of small field output factors at different field sizes were achieved independent of the jaw and MLC position. A mean and maximum pRE of 0.15% and 0.80%, respectively, were found independent of contribution from the leaf-end transmission. Augmentation of the data improved the pRE by 10%. The mean pRE for trained linear models with and without regularization was 4.23% and 9.93%, respectively.
Conclusion: We propose a machine learning-based approach for fast and accurate predictions of small field output factors. The method negates the need for further time consuming and complicated measurements without affecting the accuracy of the data. The predictions can serve as a basis for dose calculations for increased accuracy and safety of patient treatments.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
