Visual grading regression: analysing data from visual grading experiments with regression models

Abstract

For visual grading experiments, which are an easy and increasingly popular way of studying image quality, hitherto used data analysis methods are often inadequate. Visual grading analysis makes assumptions that are not statistically appropriate for ordinal data, and visual grading characteristic curves are difficult to apply in more complex experimental designs. The approach proposed in this paper, visual grading regression (VGR), consists of an established statistical technique, ordinal logistic regression, applied to data from single-image and image-pair experiments with visual grading scores selected on an ordinal scale. The approach is applicable for situations in which, for example, the effects of the choice of imaging equipment and post-processing method are to be studied simultaneously, while controlling for potentially confounding variables such as patient and observer identity. The analysis can be performed with standard statistical software packages using straightforward coding of the data. We conclude that the proposed statistical technique is useful in a wide range of visual grading studies.