Room: Exhibit Hall
Purpose: The authors have previously shown that neural network (NN) can successfully perform photon beam profile deconvolution--the elimination of volume average effect (VAE) of scanning ionization chambers (IC). The purpose of this work was to evaluate the robustness of the method when applied to ICs of various sizes and beams of different modalities (FF & FFF).
Methods: The proposed NN has an input, hidden and output layer. It inputs data extracted from cross beam profiles using a sliding window, and outputs deconvolved data at the center of the window. Cross beam profiles of fields ranging from 2x2 to 10x10 cm2Â¬ were measured with CC04, CC13, Farmer chamber and an EDGE diode detector for 6 MV FF and FFF beams. The profiles measured with each chamber were divided into training, validation, and testing sets to train and test a 3-layer feed forward NN. The diode-measured data was used as the reference. The sliding window length and the number of hidden neurons were optimized such that the same NN structure could be applied for all tested chambers, fields, depths, and beam modalities. The NNâ€™s performances were quantified by evaluating mean square error (MSE) and penumbra width difference (PWD) between the deconvolved and EDGE-measured profiles.
Results: Excellent agreement between the deconvolved and reference profiles was achieved using a sliding window width of 15 and 5 hidden neurons for all the tested ICs and both beam modalities. The average PWD decreased from 2.70Â±0.47, 2.66Â±0.41, and 3.99Â±0.42 mm to 0.09Â±0.39, 0.03Â±0.35, and 0.04Â±0.38 mm for the CC04, CC13, and Farmer chambers, respectively.
Conclusion: We found that the NN-based deconvolution method can be effectively applied to ICs of various sizes and beams of different modalities. Separate NNs are needed for different ICs but, for a specific IC, one NN works for both beam modalities.