Wavelet-informed deep video denoising for Cherenkov imaging of radiation therapy

Cherenkov imaging provides real-time video of beam incidence upon the patient, for verification of safe and accurate radiotherapy delivery. However, the optical signal is inherently weak and is affected by non-optical radiation leakage and stray x-ray noise from the medical linear accelerator (Linac). This frequently leads to low signal-to-noise ratio (SNR) frames with background clutter, limiting video image clarity and beam visualization. To address this challenge, a wavelet-based deep video denoising method was proposed. The method was validated with three regular square fields and clinical data from two volumetric modulated arc therapy (VMAT) fractions administered to breast cancer patients. Image quality was assessed using the global gamma pass rate (γ_pass) with 3%/3 mm criteria. The decision-making process of the network was visualized for interpretability. Results show that Cherenkov frames accumulated over five Linac pulses achieved γ_pass of 96-97% for all square beams. For clinical VMAT cases, accumulated frames from selected control points reached γ_pass exceeding 95%. We believe this to be the first demonstration of a deep video denoising framework sufficiently fast for real-time Cherenkov imaging.