A fast and effective denoising solution using deep learning for real time X-ray Acoustic Computed Tomography

The X-ray acoustic (XA) computed tomography has recently been proposed as a method for real-time 3D in-vivo patient dosimetry for radiation therapy. The XA effect follows the same principles as the photoacoustic effect: acoustic waves are induced due to the absorption of heat energy by the tissue from a pulsed photon beam. XA signals are small in amplitude and suffer from interference from RF noise generated by the Linear Accelerator electronics. For a real time dose reconstruction, a fast and effective denoising solution is required to increase the signal to noise in the measured XA signals. Here, we present a method to denoise the XA signals using deep learning neural networks. A Convolutional Neural Network (CNN) that operates on the spectral domain of XA signals is used. Given a noisy XA spectrogram, the CNN predicts clean XA signals. An advanced numerical model for time domain propagation of XA waves (kWave) is used to generate the training data for the CNN. Theoretical and experimental clean and noisy XA signals are obtained by megavoltage energy X-rays with long pulse width (4 us) generated from a clinical linear accelerator.