A three-dimensional cross-directional bilateral filter for edge-preserving noise reduction of low-dose computed tomography images

Background: Image-based noise reduction techniques are useful because they can be applied across various computed tomography (CT) scanner models from different vendors, regardless of the iterative reconstruction availability. The purpose of this study was to propose a 3-dimensional cross-directional bilateral filter (3D-CDBF) and compare the edge-preserving noise reduction on low-dose CT images to a model-based iterative reconstruction (MBIR). Methods: The 3D-CDBF comprises a bilateral filter and a smoothing filter applied in range filtering. The filtering process was applied with four iterations using empirically determined parameters that yielded the best tradeoff between noise reduction and edge preservation for a very low radiation dose of 2.5 mGy. In-plane and z-directional edge preservation performances for low-contrast rod phantoms (60 Hounsfield units) were compared to a clinically available MBIR and a conventional 3D bilateral filter (3D-BF), using task-based spatial resolution (task-based transfer function: TTF) and slice thickness. Moreover, the noise power spectra (NPS) were compared. Furthermore, performance was compared on abdominal CT images acquired from volunteers at 2.5 mGy (approved by our institutional review board). Results: In phantom tests, 3D-CDBF provided 28.5% higher spatial resolution at 50%TTF compared to MBIR. Moreover, total NPS was lower, while the slice thickness (z-axis resolution) was slightly broader than that achieved by MBIR (0.99 mm vs. 0.92 mm). 3D-BF was inferior to both 3D-CDBF and MBIR in all measurements. Consistent with phantom results, 3D-CDBF significantly reduced noise on abdominal images compared to MBIR (P < 0.001), exhibiting better preservation of organ edges. Conclusion: This 3D-CDBF may provide superior edge preserving noise reduction of low-dose CT images compared to currently available MBIR.