Denoising for High Fidelity External Beam Therapy Video Cherenkov Images Using a Deep Learning Methodology
Abstract
Purpose
Cherenkov Imaging is the method by which the light emitted from a patient’s skin during external beam radiation therapy can be imaged and used to verify dose delivery location. This is accomplished using an intensified CMOS camera time-gated to the linac pulse structure. The output images are heavily filtered to reduce noise; however, despite this processing, single-frame images experience excessive Poisson, Salt-and-Pepper, and Dark current noise. Additionally, the Poisson noise is blurred due to the intensifier electronics, creating a unique noise profile. This study provides updates on ongoing research to improve the SNR of Cherenkov video imaging.
Methods
Various deep learning methods have been investigated, including diffusion denoising using sub-cumulative frames as different noise levels to incorporate real noise instead of pure Gaussian, a wavelet-based method, as well as a simpler U-Net based denoiser. All models were trained using Cherenkov data retrospectively selected from the Cheshire Medical Center in Keene, NH. The cumulative images of each treatment were used as the ground truth, while single-frame and sub cumulative images acted as input data. Regarding the diffusion-based method, inference was conducted with varying denoising steps to optimize single-frame denoising. Model performance was assessed using the peak signal-to-noise ratio (PSNR) and the structural similarity index (SSIM).
Results
Regarding the diffusion method, the maximum PSNR and SSIM were found to exist between 5 and 20 denoising inference steps. The PSNR saw a ~25% increase with the SSIM increasing by ~225% for the test data displayed in figure 3.
Conclusion
While diffusion denoising provides promising results, the inference time is slower than other approaches; therefore, future work will continue to improve the computational efficiency of this model as well as develop robust denoising models, such as a wavelet or U-Net based approach with less computational cost.