Feasibility of photon beam profile deconvolution using a neural network. Issue 12 (25th October 2018)

• Record Type: Journal Article
• Authors: Liu, Han; Li, Feifei; Park, Jiyeon; Lebron, Sharon; Wu, Jian; Lu, Bo; Li, Jonathan G.; Liu, Chihray; Yan, Guanghua
• Is Part Of: Medical physics. Volume 45:Issue 12(2018)
• Page Start: 5586
• Page End: 5596
• Publication Date: 2018-10-25
• Subjects: deconvolution -- neural network -- volume averaging effect; Medical physics -- Periodicals; Medical physics; Geneeskunde; Natuurkunde; Toepassingen; Biophysics; Periodicals; Periodicals; Electronic journals; 610.153
• DOI: 10.1002/mp.13230
• Languages: English
• ISSNs: 0094-2405

Abstract

Abstract : Purpose: Ionization chambers are the detectors of choice for photon beam profile scanning. However, they introduce significant volume averaging effect (VAE) that can artificially broaden the penumbra width by 2–3 mm. The purpose of this study was to examine the feasibility of photon beam profile deconvolution (the elimination of VAE from ionization chamber‐measured beam profiles) using a three‐layer feedforward neural network. Methods: Transverse beam profiles of photon fields between 2 × 2 and 10 × 10 cm 2 were collected with both a CC13 ionization chamber and an EDGE diode detector on an Elekta Versa HD accelerator. These profiles were divided into three datasets (training, validation and test) to train and test a three‐layer feedforward neural network. A sliding window was used to extract input data from the CC13‐measured profiles. The neural network produced the deconvolved value at the center of the sliding window. The full deconvolved profile was obtained after the sliding window was moved over the measured profile from end to end. The EDGE‐measured beam profiles were used as reference for the training, validation, and test. The number of input neurons, which equals the sliding window width, and the number of hidden neurons were optimized with a parametric sweeping method. A total of 135 neural networks were fully trained with the Levenberg–Marquardt backpropagation algorithm. The one with the best overall performance on the training and validation dataset was selected to test its generalization ability on the test dataset. The agreement between the neural network‐deconvolved profiles and the EDGE‐measured profiles was evaluated with two metrics: mean squared error (MSE) and penumbra width difference (PWD). Results: Based on the two‐dimensional MSE plots, the optimal combination of sliding window width of 15 and 5 hidden neurons was selected for the final neural network. Excellent agreement was achieved between the neural network‐deconvolved profiles and the reference profiles in all three datasets. After deconvolution, the mean PWD reduced from 2.43 ± 0.26, 2.44 ± 0.36, and 2.46 ± 0.29 mm to 0.15 ± 0.15, 0.04 ± 0.03, and 0.14 ± 0.09 mm for the training, validation, and test dataset, respectively. Conclusions: We demonstrated the feasibility of photon beam profile deconvolution with a feedforward neural network in this work. The beam profiles deconvolved with a three‐layer neural network had excellent agreement with diode‐measured profiles.