Key points are not available for this paper at this time.
The assessment of image quality in medical imaging often requires observers to rate images for some metric or detectability task. These subjective results are used in optimization, radiation dose reduction or system comparison studies and may be compared to objective measures from a computer vision algorithm performing the same task. One popular scoring approach is to use a Likert scale, then assign consecutive numbers to the categories. The mean of these response values is then taken and used for comparison with the objective or second subjective response. Agreement is often assessed using correlation coefficients. We highlight a number of weaknesses in this common approach, including inappropriate analyses of ordinal data and the inability to properly account for correlations caused by repeated images or observers. We suggest alternative data collection and analysis techniques such as amendments to the scale and multilevel proportional odds models. We detail the suitability of each approach depending upon the data structure and demonstrate each method using a medical imaging example. Whilst others have raised some of these issues, we evaluated the entire study from data collection to analysis, suggested sources for software and further reading, and provided a checklist plus flowchart for use with any ordinal data. We hope that raised awareness of the limitations of the current approaches will encourage greater method consideration and the utilization of a more appropriate analysis. More accurate comparisons between measures in medical imaging will lead to a more robust contribution to the imaging literature and ultimately improved patient care.
Keeble et al. (Tue,) studied this question.