Penalty driven enhanced self-supervised learning (Noise2Void) for CBCT denoising

Abstract

Self-supervised learning for CT image denoising is a promising technique because it does not require clean target data that are usually unavailable in the clinic. Noise2void (N2V) is one of the famous methods to denoise the image without paired target data and it has been used to denoise optical images and also medical images such as MRI, and CT. However, the performance of the N2V is still limited due to the restricted receptive field of the network and it decreases the prediction performance for CT images that have complex image context and non-uniform Poisson random noise. Thus, we proposed enhanced N2V that utilizes penalty-driven network optimization to further denoise the images while preserving the important details. We used the total variation term to further denoise the image and also the laplacian pyramids term to preserve the important edges of the image. The degree of the influence of each penalty term is controlled by the hyperparameter value and they are optimized to achieve the best image quality in terms of noise level and structure sharpness. For the experiment, the real dental CBCT projection data were used to train the network in the projection domain. After the network training, the test results were reconstructed and compared at each different dose level. Meanwhile, PSNR, SNR, and a line profile were also evaluated to quantitatively compare the original FDK images, and proposed method. In conclusion, the proposed method achieved further denoises the image than N2V even preserving the details. By penalty-driven optimization, the network was able to learn the spectral features of the image while still the receptive field is limited to avoid identity mapping. We hope that our method would increase the practical utility of network-based CT images denoising that usually the target data are unavailable.