MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

A Method for TravelTime Extraction From Thermoacoustic Signal in Range Verification Based On Neural Network Algorithm

D Zhang*, X Zhuo, H Peng, S He, Y Yang, School of Physics and Technology, Wuhan University, Wuhan, Hebei, China

Presentations

(Wednesday, 7/17/2019) 9:30 AM - 10:00 AM

Room: Exhibit Hall | Forum 7

Purpose: To estimate traveltime of signal emitted from Bragg peak (γ-wave) automatically and accurately from complex thermoacoustic waveform that is comprised of γ-wave and wave emitted from proton path (α-wave), for the purpose of real time in vivo proton range verification.

Methods: A U-shaped neural network model was designed and trained based on a heterogeneous model simulating the complexity of real internal tissue and organs. Several signal receivers were put around the model. The thermoacoustic waveforms were simulated by k-Wave toolbox, in which delivery range of proton was controlled by initial energy and incidence position of proton beam. The proton beam was generated with 1.1775μs FWHM of temporal Gaussian pulse and 5 mm FWHM of radial dose distribution, and the initial energy is between 55-220 MeV. The training labels of traveltime were extracted manually from the waveforms generated by a Dirac point source located at the same position of Bragg peak. Training and testing datasets comprised 240,000 and 60,000 simulated waveforms respectively in which each waveform was sampled into 6001 points with 50 ns time interval.

Results: The results showed that the overall relative error of extracted γ-wave traveltime was less than 1% under noise-free condition. When white Gaussian noise was added to waveforms, relative error was still less than 2%, which demonstrated effectiveness of the proposed method, particularly for the case of α-wave and γ-wave aliasing.

Conclusion: The aliasing of α-wave and γ-wave brings difficulties to manual extraction of γ-wave traveltime that depends on the positions of compression peak and sparse peak. Thanks to a large number of training data extracted from comprehensively simulated waveforms, the trained neural network can extract γ-wave traveltime automatically and effectively. This deep learning algorithm for γ-wave extraction has prospect application to provide more reliable basis for thermoacoustic proton range verification.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email