Standard and Reduced Radiation Dose Liver CT Images: Adaptive Statistical Iterative Reconstruction versus Model-based Iterative Reconstruction—Comparison of Findings and Image Quality

Abstract

Liver CT images reconstructed with MBIR model-based iterative reconstruction may allow up to 59% radiation dose reduction compared with the dose with ASIR adaptive statistical iterative reconstruction , without compromising depiction of findings or image quality.

Vessel Segmentation for Noisy CT Data with Quality Measure Based on Single-Point Contrast-to-Noise Ratio

“…For example, CT reconstruction quality for different radiation doses is investigated using phantom images [9,24]. These metrics measure different aspects of image quality: PSNR describes degradation of the best signal, while the method noise measures image edge corruption by denoising procedures.…”

“…Usually, it is necessary to choose one ROI on the object and one -on the background to compute and [13,24]. However, it is possible to get incorrect CNR estimation on nonuniform image parts [14].…”

“…In the simple twopoint method we use one point inside and one outside of the vessel object. Similar to ROI selection in [24], a two-point CNR has the following form: (2) where и are points at the vessel and surrounding tissue, respectively, is -th component of , is an intensity value at the point , M is an intensity median over a cubic neighborhood, is the size of the cubic neighborhood on the vessel (object), -on the surrounding tissues (background), is defined the same way as , is standard deviation across the same region as , is or Chebyshev distance. The computation of can be done assuming unimodality of the intensity distribution in the cubic image patch centered in .…”

“…Multiple methods exist to perform image restoration both at the scanning and reconstruction stages [24], and at the image processing stage [3].…”

“…However, there is no unambiguous solution to measure image quality in practical CT segmentation tasks, because to use a common PSNR metric we would need a noise-free image. Contrast-to-noise measures require a ROI in the image to be selected [24]. This prior knowledge is available only for synthetic tests or when we already have "ground truth" segmentation.…”