(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose:
Conventional 2D electron treatment requires cutout factor measurement for MU calculation. The purpose of this study is to test the feasibility of using deep learning (DL) to estimate electron cutout factor without measurement or Monte Carlo simulation.
Methods:
Eclipse eMC (v15.5) was clinically validated in our institution for profiles and output factors with the error tolerance of 1% and 3%, respectively. To collect training data, eMC calculations were performed on 145 cutouts with various sizes and shapes including circle, square, rectangle, triangle, and L-shapes. A fixed machine type (Varian Clinac), energy (9 MeV), electron cone (20x20 cm²) and SSD (105 cm) were used in this study. Eclipse scripting API and a Matlab code was used to export output factors with corresponding cutout polygons and to populate randomly rotated and translated cutout images. Using 2611 training data, a CNN-regression model was built with 4x4 convolutional layers, each followed by a rectified linear unit, and then one max pooling operation was used before fully connected layers. A batch normalization and dropout layers were further added to each convolutional layer. For validation, the cutout factors of 38 clinical cases with the minimum width of 3 cm were calculated using the DL model and then compared to those from eMC.
Results:
The mean absolute error of the DL-based cutout factors compared to the eMC-based results was 0.8±0.6%. The average calculation time of the DL-based method was <0.2 s, compared to >5 mins using eMC calculation (with 8 processors) and >30 mins using manual measurement.
Conclusion:
Our results showed that DL-based cutout factor estimation is feasible and it can be used as an alternative method to MC simulation or direct measurement for 2D electron treatment planning. Future work includes creating a general DL model that can account for various electron energies, applicators and SSDs.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
