Feasibility study of range verification based on proton-induced acoustic signals and recurrent neural network

Abstract

Range verification in proton therapy is a critical quality assurance task. We studied the feasibility of online range verification based on proton-induced acoustic signals, using a bidirectional long-short-term-memory recurrent neural network and various signal processing techniques. Dose distribution of 1D pencil proton beams inside a CT image-based phantom was analytically calculated. The propagation of acoustic signal inside the phantom was modeled using the k-Wave toolbox. For signal processing, five methods were investigated: down-sampling (DS), DS + HT (Hilbert transform), Wavelet decomposition (Wavedec db1, db4 and db20). The performances were quantitatively evaluated in terms of mean absolute error, mean relative error (MRE) and the Bragg peak localization error ( Δ B P ). In addition, the study analyzed the impact of noise levels, the number of sensors, as well as the location of sensors. For the noiseless case (32 sensors), the Wavedec db1 method demonstrates the best performance: Δ B P is less than one pixel and the dose accuracy over the region adjacent to the Bragg peak (MRE50) is ∼3.04%. With the presence of noise, the Wavedec db1 method demonstrates the best noise immunity, achieving Δ B P less than 1 mm and an MRE50 of ∼12%. The proposed machine learning framework may become a useful tool allowing for online range verification in proton therapy.

Feasibility study of 3D time‐reversal reconstruction of proton‐induced acoustic signals for dose verification in the head and the liver: A simulation study

“…To elucidate, by selecting only the 16 sensors within a single slice (e.g., liver case, slice #185, same sensor arrangement, same acoustic signals obtained from 3D forward propagation), the results of 3D TR and 2D TR are obtained (see Figures S1 and S2 in the supporting information). We find that given the small number of sensors (16), both show inferior results and thus make it difficult to draw any definitive conclusion. For the time being, we are not ready to answer whether one would obtain the comparable results as those in Figure 3 (e.g., 208 sensors evenly distributed in multiple slices), when putting 208 sensors densely packed within a single slice (2D).…”

“…Compared to the TOF‐based method, a TR approach is believed to be more flexible, allowing for arbitrary detection geometry and tolerating tissue heterogeneity. The second approach is a machine learning‐based approach using recurrent neural networks, which has great potential to provide an end‐to‐end verification framework with profoundly reduced computational workload 16 …”

“…The second approach is a machine learning-based approach using recurrent neural networks, which has great potential to provide an end-to-end verification framework with profoundly reduced computational workload. 16 The results of two-dimensional (2D) TR study have been reported in our previous work, 15 which comprehensively examined the impact of six parameters, including the number of sensors, sampling duration, sampling timestep, spill time, noise level, and number of iterations. To move a step closer towards practical applications, this paper extends the TR approach from 2D to 3D.…”

Automatic detection and classification of treatment deviations in proton therapy from realistically simulated prompt gamma imaging data

“…In the field of treatment verification, ML has already been used to convert simulated distributions of PGrays, 24 positron emitters, [25][26][27] and acoustic signals 28,29 to dose distributions. The mentioned publications use simulations on phantom data and mostly only investigate high-weighted PBS spots.…”

Deep learning-based protoacoustic signal denoising for proton range verification

“…In previous studies, denoising techniques such as low-pass filters [16], wavelet-based transformations (WT) [17,18] and correlation based methods [19] were used to eliminate high-frequency noise in signals.…”