Deep learning-based protoacoustic signal denoising for proton range verification.

Area of Science:

• Medical Physics
• Radiation Oncology
• Artificial Intelligence

Background:

• Proton therapy offers superior dose distribution via the Bragg peak (BP).
• Current protoacoustic techniques for in vivo BP localization require high radiation doses for adequate signal-to-noise ratio (SNR).

Purpose of the Study:

• To develop and evaluate a deep learning-based denoising technique for protoacoustic signals.
• To reduce the necessary radiation dose and improve BP range uncertainty in proton therapy verification.

Main Methods:

• Utilized three accelerometers to collect protoacoustic signals from a polyethylene phantom.
• Trained device-specific stack autoencoder (SAE) models for denoising low number-of-signal-averaging (NSA) data.
• Compared supervised and unsupervised SAE training strategies using metrics like mean squared error (MSE), SNR, and BP range uncertainty.

Main Results:

• Supervised SAE models demonstrated superior performance in BP range verification compared to unsupervised models.
• Achieved low BP range uncertainty (e.g., 0.20 ± 3.44 mm with 8 NSA) using the deep learning approach.
• Significantly enhanced SNR and reduced uncertainty even with substantially lower NSA, indicating reduced dose requirements.

Conclusions:

• Deep learning-based denoising effectively enhances protoacoustic signal quality.
• This method shows promise for accurate in vivo BP range verification in clinical proton therapy.
• Reduced dose and time requirements make the technique suitable for clinical translation.