Radiation-induced Acoustic Signal Denoising using a Supervised Deep Learning Framework for Imaging and Therapy...

Area of Science:

• Medical Imaging
• Biophysics
• Machine Learning

Background:

• Radiation-induced acoustic (RA) imaging visualizes radiation dose deposition but requires extensive averaging for adequate signal-to-noise ratios (SNRs).
• High averaging requirements increase radiation exposure and limit temporal resolution, hindering clinical applications of RA imaging.
• Existing methods struggle to balance SNR, radiation dose, and imaging speed.

Approach:

• Developed a general deep inception convolutional neural network (GDI-CNN) for denoising RA signals.
• Utilized multi-dilation convolutions to capture diverse temporal signal features, ensuring network generalizability.
• Validated the GDI-CNN on X-ray-induced acoustic, protoacoustic, and electroacoustic signals.

Key Points:

• GDI-CNN achieved comparable SNRs to fully-averaged signals using less than 2% of the averages across multiple RA modalities.
• The method substantially reduces ionizing radiation dose required for RA imaging.
• Significant improvement in temporal resolution was observed, enabling real-time monitoring.

Conclusions:

• The GDI-CNN framework offers a general solution for few-frame-averaged acoustic signal denoising.
• This deep learning approach enhances the clinical utility of RA imaging for low-dose applications.
• The technology supports real-time monitoring of therapeutic interventions using RA imaging.