Improved Estimation of Tissue Noise Power Spectra in CT Data

Evaluation and measuring of image quality in X-ray computed tomographic (CT) data gained importance with recent appearance of modern algorithms for iterative reconstruction of CT data. Thanks to the ability of dramatically reducing applied radiation dose declaratively without loss of image quality, they are expected to replace the conventionally used filtered back projection (FBP) algorithm. Quality of iteratively reconstructed data in terms of image noise is routinely evaluated in images of homogeneous phantoms or in small regions of interest in real patient data. Character of the noise, whose characteristics are dependent on imaged scene, require measuring in the whole volume of real patient data and moreover in diverse tissues separately. This paper presents generalization of one dimensional noise power spectra estimation which enables its calculation from separate tissues. Firstly, basic tissues must be segmented and the resulting segmentation masks are used for the noise power spectra estimation. The estimation carried out with the help of the binary segmentation masks is, due to convolutional property of the Fourier transform, burdened by error due to spectral leakage. A binary segmentation mask may be seen as a two-dimensional windowing function with steep borders. Our method for reduction of the error is based on replacement of binary segmentation masks by designed two-dimensional spatially adaptive windowing functions with better spectral properties. Design of the spatially adaptive windows is based on distance maps and optimized skeletonization calculated using the maximal discs approach. The magnitude of the segmentation introduced error can be experimentally measured using a simulated noise with known power spectrum, which is compared with the noise power spectrum estimated in frame of the segmented tissue (i.e. affected by the spectral leakage). Finally, it is shown that the proposed two-dimensional spatially adaptive windowing functions are able to significantly improve precision of the noise power spectra estimation in diverse tissues