2.5D Extension of Neighborhood Filters for Noise Reduction in 3D Medical CT Images

摘要

Noise in 3D computer tomography (CT) images is close to white and becomes large when patient radiation doses are reduced. We propose two methods for noise reduction in CT images: 3D extension of fast rank algorithms (Rank-2.5D) and 3D extension of a non-local means algorithm (NLM-2.5D). We call both our algorithms "2.5D" because the extended NLM algorithm is slightly asymmetric by slice axes, while our Rank algorithms, being fully symmetric mathematically and by results, have some implementation asymmetry. A comparison of the methods is presented. It is shown that NLM-2.5D method has the best quality, but is computationally expensive: its complexity quickly rises as a function of the neighborhood size, while Rank-2.5D only shows linear growth. Another contribution of this paper is a modified multiscale histogram representation in memory with a tree-like structure. This dramatically reduces memory requirements and makes it possible to process 16-bit DICOM data with full accuracy. Artificial test sequences are used for signal-to-noise performance measurements, while real CT scans are used for visual assessment of results. We also propose two new measures for no-reference denoising quality assessment based on the autocorrelation coefficient and entropy: both measures analyze randomness of the difference between noisy and filtered images.