Diagnostic medical imaging is an essential element of modern health care systems. Health care systems should provide the level of image quality that allows radiologists to accurately identify pathology without errors. The assessment of image quality for medical imaging technology is generally based on technical measurements (e.g. transfer functions) by physicists. The assessment of clinical performance (clinical image quality) is generally based on clinical studies and performed by radiologists.
There are many tasks in radiology departments that involve assessment of image quality. For instance, tests to approve a system for clinical use or tests to monitor its technical stability over time. Such tests are mainly performed by objective measurements of physical characteristics. Basically these kinds of measures express physical image quality aspects as contrast, sharpness and image noise. The advantage of physical measurements is that they are well described, reproducible and relatively easy and fast to perform. Alternatively, to find best ways to use the imaging system for specific clinical purposes, anthropomorphic phantom studies and clinical studies are used for subjective (quality rating) and objective (lesion detection) observer performance studies. Performing clinical studies with human observers is time consuming and costly because they require a large number of human observations and a significant number of observers. Moreover, the number of possible conditions to investigate can be extremely large.
The detective Quantum Efficiency (DQE) is a measure of the combined effect of the noise and contrast performance of an imaging system, it is expressed as a function of spatial frequency. Noise can be expressed by the signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR) or by the noise power spectrum (NPS). An imaging system’s ability to render the contrast of an object as a function of spatial frequency is traditionally expressed as its modulation-transfer function (MTF). The combination of the functions NPS and MTF determines the above mentioned DQE. These objectives physical measurements describe the systems technical imaging performance but it is still difficult to translate the outcome to the clinical situation that is more complex than these measurements can describe.
The most fundamental factors in physical image quality are contrast, sharpness and noise. A semi-objective physical test that is often used to measure the combined effect of these factors is ‘low contrast detectability’ (LCD; also ‘contrast-detail detectability’). This test is more directly related to the clinical detection task than the objective physical measurements above. Low contrast detectability is most often determined by human observers by scoring the visibility of low contrast objects within phantom images. The phantoms used are built from homogeneous material (e.g. PMMA) and contain low contrast signals. The homogeneous material aims to mimic the patient mainly just by scattering, filtering and attenuating radiation. This method might be biased as the low contrast objects in the phantom are arranged in fixed patterns the observer may know beforehand (reducing objectivity). Besides, inter and intra-observer variability may exist and analysis is time-consuming. A widely used alternative for humans to evaluate detectability in medical images objectively are model observers – that are numerical methods acting as surrogates of the human performance (find more detail on model observers below). For instance, to more clearly demonstrate the impact of parameter settings and new technology on the CT image quality we developed an automated and objective method to investigate low contrast detectability using a model observer.
One should be careful in directly translating the low contrast detectability results to the clinical situation, by concluding what combination of object size and contrast would be visible in clinical Images. The objects in LCD studies are often cylinder shaped showing sharp borders in contrast to clinical lesions. Moreover, the appearance of the objects in the final homogeneous images may be different to what it would have been in actual processed patient images with anatomical background. The above measurement is therefore often intended for characterizing the technical performance of the image receptor in combination with the x-ray tube.
Human observer studies may become very time consuming and costly because they require a large number of human observations and a significant number of observers. Moreover, the number of possible conditions that are of interest to investigate can be so large, making human observation studies practically impossible. As an alternative to human observers, a computer-model of human observers can be considered. Such an algorithm predicts human visual performance in diagnostic images. The tasks to be performed by the model observer can be divided generically into classification and estimation tasks. In medical applications, an example of a classification task would be lesion detection (for instance it can estimate human detection probability of a certain object), while an estimation task might be determination of the volume of blood expelled from the heart on each beat. The model observer can be used for system evaluation and optimization with the assumption that the system that is best for the model is also best for a human. These models seem very useful in investigating numerous different conditions in diagnostic medical imaging. More specifically, it is widely accepted that specific model observers (e.g. Channelized Hotelling Observer and NPWE model observer) correlate well with the human observer.