Encrypted login | home

Program Information

Optimization of An Artificial Neural Network for Photon Beam Profile Deconvolution

F Li

F Li*, S Lebron , J Wu , B Barraclough , J Park , C Liu , G Yan , University of Florida, Gainesville, FL


TU-FG-205-12 (Tuesday, August 1, 2017) 1:45 PM - 3:45 PM Room: 205

Purpose: Volume averaging effect (VAE) of finite-sized ion chambers (IC) negatively affects the penumbra of measured photon beam profiles, leading to incorrect beam models. The removal of VAE via numerical deconvolution had little success due to various reasons (e.g., noise). This work seeks to obtain a robust, accurate and direct solution by optimizing an artificial neural network (ANN).

Methods: In Fourier domain, the challenging deconvolution problem turns into a convolution problem, lending itself naturally to the use of ANN. A fully connected 1- or 2-layer ANN was trained to recover the “true” beam profiles from IC measurement. A sliding window was used to extract input from the measurement. The hidden and output layer of the ANN used tanh and linear activation function, respectively. The ANN output the deconvolved value at the center of the sliding window. The backpropagation algorithm with stochastic gradient descent was used to train the ANN where diode-measured beam profiles were used as desired output. We evaluated the performance of the ANN by varying (1) the spread of the sliding window (2) the number of hidden layers (3) the number of neurons and (4) the noise level. We also studied its performance with various detectors (CC13, CC04 and SNC125).

Results: The ANN achieved superior accuracy for all three detectors with RMSE<0.002 on training data and RMSE<0.003 on validation data. Overall, best performance was achieved when the size of the sliding window was 9 and the number of hidden neurons was 7. The additional hidden layer did not improve the accuracy significantly.

Conclusion: Accurate and robust solutions to directly remove IC VAE in photon beam profiles were obtained using ANN for all studied detectors. Since the IC VAE is mainly a function of detector size, these well-trained ANN can be readily incorporated into commercial scanning software.

Contact Email: