A comparison study on KL domain penalized weighted least-squares approach to noise reduction for low-dose cone-beam CT

Dose reduction is a major task for cone-beam computed tomography (CBCT) applications because of the potential side effect of X-ray exposure to the patients. One of the strategies to achieve low-dose is to lower the X-ray tube current and/or shorten the exposure time in CT scanners. However, the image quality from the low mAs acquisition is severely degraded due to excessive quantum noise. In this work, we investigated three implementations of Karhunen-Loéve domain penalized weighted least-squares (KL-PWLS) scheme to adaptively treat the noise in the low-mAs CBCT sinograms. The motivation is based on the observations that strong data correlation exists between neighboring views and neighboring slices in CBCT and the KL transform can de-compose the correlated signals for adaptive noise treatment. The three implementations were: (1) performing the KL transform among neighboring views and reduce the three-dimensional (3D) noise-treatment procedure into a series of 2D operations, (2) performing the KL transform among neighboring slices in the 3D space, and (3) performing the KL transform on a view-by-view manner along the slice direction for sparse data sampling which actually reduces the procedure into a series of 1D operations. The noise-treated sinogram data were then reconstructed by the analytical Feldkamp-Davis-Kress (FDK) algorithm. The effectiveness of the presented KL-PWLS noise reduction strategy was evaluated using two physical phantoms (CatPhan600® and anthropomorphic head). Noise in the reconstructed CBCT-FDK images from a low 10mA protocol was greatly suppressed without noticeable sacrifice of the spatial resolution compared with those from a high 80mA protocol, which implies a potential dose reduction by as high as a factor of 8 for the two phantoms. The noise-resolution tradeoff curves indicate that the KL-PWLS implementations considering the neighboring slices (implementation 2 and 3) outperform that considering the neighboring views (implementation 1) in better resolution preservation at the same noise level, which is probably due to the structural continuity among neighboring slices. For further dose reduction by sparse data sampling, implementation (3) can be a potential choice attributed to its computational advantages over the other two implementations.