Deep Learning Sinogram Denoising for Low-Dose CT Image Quality

A deep learning neural network is introduced as an intermediary component between the filtering step and the analytical reconstruction step. The neural network processes the filtered sinogram data, denoising it and preserving high-frequency components before passing it to the back-projection algorithm. This mediator enables the system to maintain both fast reconstruction speed and high image quality by preprocessing the data in a way that compensates for the limitations of filtered back-projection in low-dose scenarios

The patent converts the harmful effect of high quanta noise in low-dose images into a beneficial outcome by using the neural network to identify and preserve meaningful high-frequency signal components while removing noise. The neural network learns from training data to distinguish between noise and actual image features, effectively transforming the low-dose constraint into an opportunity to develop more sophisticated noise reduction capabilities that improve image quality without requiring higher radiation doses

The patent applies local quality by treating different frequency components of the sinogram data differently through the neural network. The network selectively processes high-frequency components with special attention, applying denoising operations that are adaptive to the local characteristics of the data. This localized processing approach ensures that important high-frequency information is preserved while noise is suppressed, improving reconstruction performance without requiring complete redesign of the entire filtering process