(Saturday, 4/7/2018)
Room: Foyer
Purpose: To deconvolve volume averaging effect (VAE) of ion chambers (IC) in beam profile measurement using neural network (NN) and to investigate whether separate NNs for flattening filter (FF) fields and flattening filter free (FFF) fields are necessary.
Methods: A feed-forward NN with a hidden layer and an output layer was trained with profiles measured with an IC and a diode. The hidden and output layer used tan-sigmoid and linear transfer function, respectively. A sliding window was used to extract input data for the NN from the IC-measured profiles. The diode measurement at the center of each sliding window was used as the desired output. The NN, trained with a Levenberg Marquardt-based backpropagation algorithm, output a deconvolved value for each sliding window. We optimized the size of the sliding window (SSW) and the number of hidden neurons (NHN) as hyperparameters. Separate NNs were trained for FF and FFF fields, respectively. Combined NNs were also trained with both FF and FFF profiles. Their performance was evaluated by the penumbra width difference (PWD) between the predicted and diode-measured profiles.
Results: For the separate NN for FF fields, minimal mean PWD of 0.03 mm was achieved with SSW= 9 and NHN=7; for FFF fields, minimal mean PWD was 0.02 mm when SSW=7 and NHN=7. The performance of the combined NN was similar with the minimal mean PWD for FF and FFF fields being 0.05 and 0.03 mm, respectively, when SSW=9 and NHN=7. Separate NNs predicted slightly smoother beam profiles than the combined NNs for small FFF fields (2x2 cm2).
Conclusion: NNs are accurate and efficient tools to minimize VAE of IC in beam profile measurement for both FF and FFF fields. It is not necessary to train separate NNs for FF and FFF beam fields.
Contact Email
