Room: Track 3
Purpose: Proton acoustic can allow for proton range verification with submillimeter accuracy in a homogeneous medium. However, it requires to deliver a large number of pulses (e.g. 512 proton pulses, which corresponds to 24.2 Gy) and a long beam delivery time (~10 minutes) to a single spot. In this study, we propose to develop a deep-learning-based signal denoising algorithm for the proton-acoustic signals using stacked auto-encoder (SAE).
Methods: For acoustic data acquisition, an accelerometer array was used on the distal surface of a cylindrical polyethylene phantom (diameter=20.88 cm, length=33.58 cm). The proton energy was 226 MeV, and the beam was attenuated by 2 cm solid water before entering the PE phantom. Sequentially, we classified the measured signals for the 400 datasets as training and 112 datasets as testing from a total of 512 raw signals. Among the datasets, we tested a various number of ground-truth from 114 to 256, and the number of averaged signals from 1 to 16. The SAE was established by stacking a series of the auto-encoder and decoder layers, and the output of the bottom layer is the input of the second layer. Both mean squared error (MSE) and signal-to-noise ratio (SNR) was used for comparison between pre-training and after-training results.
Results: Overall, the average MSE is significantly improved from 0.02405 to 0.00068, and the SNR is also enhanced from 10.89 dB to 36.84 dB with the proposed method. Even within the noisiest signal (i.e. 144 ground-truth datasets averaged over 16 signals), the proposed method shows robust enhancement with the MSE from 0.04693 to 0.00068 and the SNR from 7.25 dB to 36.85 dB.
Conclusion: We have developed a noise reduction method for proton-acoustic signals using a SAE technique. With this method, we can reduce both the delivered dose and delivery time for proton range verification.