Frequency-aware denoising using a diffusion model for enhanced band-limited and white noise removal in x-ray acoustic...

Radiation therapy delivers precise doses to tumors, but accurately measuring internal tissue doses remains a challenge. Current methods, such as ionization chambers and radiographic films, rely on external measurements, which cannot provide direct, in vivo dose feedback. X-ray acoustic computed tomography (XACT) was developed to generate thermoacoustic signals when x-rays deposit energy into water or tissue, enabling the reconstruction of dose distribution patterns through acoustic signals. However, the longer pulse width of x-rays from linear accelerators reduces the efficiency of thermoacoustic signal conversion, lowering the signal-to-noise ratio (SNR) of radiofrequency (RF) signals. This noise significantly affects the quality of reconstructed XACT images. Overcoming the impact of noise is essential for advancing XACT toward accurate dose detection.

This study aims to develop a frequency-aware denoising (FAD) method for overcoming the impact of band-limited and white noise in RF signals for XACT.

Real RF signals were acquired from an XACT system using radiotherapy megavoltage (MV) x-rays, a water tank, and an ultrasound transducer. To capture the frequency characteristics of these RF signals, we first estimated the probability density function (PDF) of their frequency spectrum. To generate synthetic RF data that closely approximates realistic noisy signals for model training, noise was sampled from this PDF, incorporating both magnitude and random phase components, and combined with simulated signals and white noise. A conditional diffusion model was trained on these synthetic signals to obtain the FAD model. A total of 3150 frequency-aware RF data samples were used to train the FAD model. For testing, acoustic RF signal data excited by five different x-ray shapes were measured, denoised by the FAD model, and finally reconstructed into XACT. The performance of the method was evaluated based on XACT image quality using SNR analysis and γ passing rate, and compared with results from Raw-RF and background noise-removed (BNR-RF) methods.

The FAD-RF method produced XACT images with clearer structural details and fewer artifacts. It achieved the highest SNR among the tested methods, with a mean SNR of 27.6 ± 5.0, outperforming both Raw-RF (22.9 ± 2.2, p < 0.05) and BNR-RF (22.0 ± 3.0, p < 0.05). In terms of spatial accuracy, the FAD-RF method also outperformed in γ analysis, achieving a mean γ passing rate of 79.0% ± 2.4%, significantly higher than Raw-RF (50.2% ± 20.0%, p < 0.05) and BNR-RF (73.5% ± 3.6%, p = 0.078).

The FAD-RF method relies solely on the RF signals for denoising, making it practical and efficient for real-world applications. It demonstrates effective noise suppression and enhanced spatial accuracy in XACT image reconstruction.