Author: Petr Bruza, Jeremy Eric Hallett, Brian W Pogue, Yucheng Tang, Shiru Wang 👨🔬
Affiliation: NVIDIA Corp, Dartmouth College, Thayer School of Engineering, Dartmouth College, University of Wisconsin-Madison, University of Wisconsin - Madison 🌍
Purpose: Cherenkov imaging allows for real-time visualization of megavoltage X-ray or electron beam delivery during radiation therapy. By using a time-gated intensified CMOS camera synchronized with a linear accelerator, Cherenkov light emitted from a patient's surface can be captured. This provides real-time display of exit and entrance dose beam maps. However, these images often suffer from high noise levels, complicating interpretation. This study focused on enhanced real-time noise suppression using a diffusion-based deep learning model. The proposed model is designed to robustly denoise both cumulative and single-frame images, producing high-quality outputs that improve the interpretability of Cherenkov imaging.
Methods: Departing from conventional diffusion models that add synthetic noise, a denoising diffusion model was developed with an innovative approach to noise management by using linear interpolation, thus naturally decreasing the signal-to-noise ratio (SNR). This method begins with the complete stack of images and gradually interpolates to a single frame, allowing for precise control over both the maximum and minimum noise levels in the input data. To train and validate the performance of the proposed model, a mixture of clinical and phantom Cherenkov images were employed.
Results: Applying the denoising model to a cumulative phantom image acquired during a 25MU irradiation resulted in a 27.2% PSNR increase when using 700 denoising steps. This noise decrease came with a minimal decrease in resolution evidenced by the 10%-MTF crossing dropping by only 5.3%. When applied to a single-frame phantom image, the PSNR increased by 5.4%, a great result for these low-signal images.
Conclusion: The proposed diffusion model exhibits superior performance when compared to traditional denoising algorithms. Furthermore, it maintains consistent efficacy across both single-frame and accumulative Cherenkov images, establishing a strong foundation for online denoising in real-time monitoring during prospective settings.