There are various papers being presented over the last years from various groups describing different approaches for determining image quality parameters for which some examples had been given in the methods section. Many of such papers are looking into computed tomography imaging nowadays which does make sense due to the relatively high exposure correlated with threedimensional imaging using ionising radiation on the one hand side and the broad use of applications on the other hand. In addition there is currently a lot of effort ongoing to improve the relation between image quality and patient exposure in CT for example by optimising hardware (like using photon counting detectors, tube current modulation or dedicated filters etc.), optimising imaging parameters (including pitch, slice thickness, beam quality etc.) or image reconstruction (iterative reconstruction, AI based reconstruction, noise reduction techniques). Thus, it seems to be reasonable to evaluate the different descriptors for image quality assessment using the example of CT imaging.
Within EUROSAFE Imaging there is a working group dealing with image quality aspects. The experts working together in this group are evaluating the advantages and drawbacks of the different approaches described above:
As already briefly mentioned subjective image quality assessment has the main advantage that it is directly related to the diagnostic performance, although such an evaluation does not necessarily rely on the detection of pathologies. If the subjective image quality evaluation is based on evaluations like the appearance of the image then it might be misleading with respect to the diagnostic performance of the method. In addition - and this is probably the most important aspect preventing the use of such image quality assessment for optimising a large set of parameters for an imaging task -, such an approach would always need a large number of readers evaluating many images of each set of parameters to generate reliable results.
Fourier-based image quality parameters derived from imaging phantoms are on the other hand pretty easy to determine. However, it can be shown, that the theoretical background for the Fourier.based image description is not met for most applications. In addition, in many studies it was hard if not impossible to show a direct relation between the image quality determined and the diagnostic performance of the imaging setups to be compared. Nowadays, with non-linear image processing the limitations of Fourier-base image quality assessment on phantom structures will be even larger. The reason for this is, that easy or regular recurrent structures will be reproduced much better than subtle non-regular structures as in real-patient images.
Task-based image quality evaluation has the advantage that - if designed properly - it can describe the detection task pretty well and thus, it can at least to a certain amount predict the diagnostic performance of an imaging setup pretty well. While this is a major advantage together with the fact that no large scale reader efforts are necessary, the workload for the generation of suitable images and the corresponding evaluation is quite large. This is especially true when taking into account, that for different diagnostic questions different tasks need to be defined and analyzed. It is not easy to estimate how task-based image quality tests would work for AI based reconstruction methods.
Taking the different advantages and drawbacks into account, various groups try to develop approaches to determine by means of physics based image quality parameters the image quality in patient images directly. Until today, it is not completely clear, whether these approaches typically based on Fourier-based image quality parameters evaluated on patient image structures can really indicate the diagnostic performance of imaging setups and image processing, but there are at least certain correlations as can be seen for example when looking at Figures 1 and 2 as taken frm 6).
Besides the limitations of the use of the Fourier-based approaches, it has also to be stated that it is not easy to define which structures should be evaluated. In addition, it can be detected that the results of some of the approaches strongly depend on some of the parameters to be optimised like e.g. the slice thickness.