Joel Larsson - Effects of nonlinear noise reduction algorithms on image quality in computed tomography systems. Evaluations using human observers and methods for assessing distortion

Link directly to the doctoral thesis

Image

This thesis underlines the importance of studying the distortion effect when nonlinear algorithms, intended to reduce visible noise, are used to reduce the radiation dose in computed tomographic examinations.

ABSTRACT

Many conventional radiological examinations have during the past decades been replaced by examinations performed on computed tomography (CT) systems. One reason is that a CT system, in contrast to a conventional X-ray system, depicts slices of the body so that anatomical structures to a lesser extent risk to obscure the potential pathology. This extra diagnostic information may increase the absorbed dose for patients, because the noise in CT examinations acquired at the same absorbed dose as conventional radiographs would have instead risked obscuring the pathology.

Hence, if the noise in the CT images could be reduced by a mathematical algorithm a reduction in absorbed dose may also be possible. Traditionally, noise is reduced using linear convolution kernels, which weights the content of the CT image such that distinct variations are reduced. Concurrently, sharp edges in the image are smoothed out (the resolution is reduced), as sharp edges and noise are described by the same type of image content. Hence, the amount of noise reduction and consequently dose reduction will be limited by the required image resolution for the diagnostic task. In contrast to a linear algorithm, a nonlinear noise reduction algorithm is intended to reduce noise while keeping or increasing the image resolution. Hence, the image quality brought about by such an algorithm may depend on the content of the image including the noise level, which will make prediction of image quality in patients more difficult than for a linear algorithm. Further, nonlinear algorithms tend to distort the image. The impression of the image content may potentially be changed and aggravate the diagnostic assessment. Thus, the overall aim of the thesis was to investigate the effects of nonlinear noise reduction algorithms in CT imaging to help understand how to assess and predict image quality.

The nonlinear effect was investigated using human observer evaluation of paediatric cerebral and abdominal CT examinations, which had been noise reduced by a nonlinear noise reduction algorithm. However, for the abdominal examination, the combination effect of type of convolution kernel and noise reduction strength on image quality was investigated. These investigations showed the visualisation of some anatomical structure to increase, concurrently the resolution of other structures was shown to decrease as the strength of the noise reduction was increased. An edge-enhanced convolution kernel showed to compensate the reduced resolution. However, the visualisation of the structures was not found to be higher than the optimal strength of the noise reduction for the original convolution kernel.

The distortion effect was investigated using an objective method implemented from conventional radiography. The method showed that the tendency to distort the image content for the nonlinear algorithms increased with the noise level. However, the method does not visualize the distortion in the spatial domain. Hence, the method inspired to develop a new method, which shows where the distortion is located and how the image is distorted when noise is reduced by a nonlinear algorithm. The new method showed the distortion of the imaged object to be caused by the noise and the distortion of the noise to be caused by the imaged object.

The thesis emphasises the importance of investigating the distortion effect when nonlinear noise reduction algorithms are used to reduce the absorbed dose of CT examinations.

The illustration above, figure 10, is from the thesis, page 58:
a) shows an example of a typical CT image with the noise reduction already applied at the start of the study
b) shows the same image but without noise reduction
c) shows the same image but with the highest possible noise reduction achieved by the investigated algorithm
d) shows the same image but with the optimal noise reduction found in the study.

Keywords: Computed Tomography, Nonlinear noise reduction, Visual grading, Distortion analysis