"External beam radiation therapy uses megavoltage photons and electrons to treat cancer while limiting healthy tissue toxicity. Cherenkov light, which is emitted from a dielectric medium when charged particles travel faster than that medium’s phase velocity, can be used to visualize treatments. This light is proportional to the radiation dose, providing a visual map of treatment. Improvements in Cherenkov imaging can be obtained using machine learning workflows. Image denoising and automatic PHI removal are two key areas for AI implementation.
Face de-identification was performed using a YOLOv7 object detection model trained on Cherenkov background images. Clinical images were obtained from two Dartmouth health centers. Various model parameters and augmentation methods were varied, including the image scaling range, batch size, final learning rate, and objective function weights, resulting in a max prediction accuracy of 96%.
In addition to image de-identification, diffusion neural network, wavelet, and Unet models were developed for Cherenkov image denoising, where the models were trained using cumulative images with a simulated Cherenkov noise model. All models showed drastic improvement compared to traditional denoising algorithms, with the diffusion model seeing a 5.53% and 74.4% increase in peak signal to noise ration and structure similarity index, respectively. This work will pave the way for enhanced Cherenkov video rate visualization of subsurface features."