To the top

Page Manager: Webmaster
Last update: 9/11/2012 3:13 PM

Tell a friend about this page
Print version

Reanalysis of visual grad… - University of Gothenburg, Sweden Till startsida
Sitemap
To content Read more about how we use cookies on gu.se

Reanalysis of visual grading characteristics (VGC) data using VGC Analyzer

Conference contribution
Authors Jonny Hansson
Lars Gunnar Månsson
Magnus Båth
Published in Optimisation in X-ray and Molecular Imaging 2015 - the Fourth Malmö Conference on Medical Imaging, Gothenburg, Sweden, 28-30 May 2015
Publication year 2015
Published at Institute of Clinical Sciences, Department of Radiation Physics
Language en
Links www.delegia.com/app/Data/ProjectIma...
https://gup.ub.gu.se/file/187456
Subject categories Radiological physics

Abstract

Purpose: Visual grading studies have a weak point in that the statistical evaluation of collected data is often performed in a questionable manner. The introduction of visual grading characteristics (VGC) analysis in 2007 aimed at an improvement by presenting a non-parametric rank-invariant method of comparing visual grading data from two modalities. The resulting figure-of-merit, the area under the VGC curve (AUCVGC), was initially proposed to be determined using standard software for receiver operating characteristics (ROC) analysis. However, due to different properties of ROC and VGC data, it can be assumed that the uncertainty of the AUCVGC is not determined correctly using ROC software. A novel developed software tool (VGC Analyzer) for statistical analysis of VGC data using non-parametric resampling methods has recently been verified on simulated data. The purpose of the present work was to reanalyse data from previously published VGC studies using VGC Analyzer in order to evaluate the validity of the reported results of the previous studies and to investigate the behaviour of VGC Analyzer on real data. Method: Visual grading data from previously performed studies on optimisation of x-ray examinations were reanalysed using VGC Analyzer. The outcome (the mean and 95% confidence interval (CI) of the AUCVGC and the p-value) was compared with previously reported data from the studies where single reader adapted ROC software had been used and rating data from multiple readers had been pooled. The studies included both paired and non-paired data and were analysed using both fixed readers and random readers. Major findings: The results showed good agreement between the AUCVGC determined with VGC Analyzer and the previously used methods. However, on non-paired data, narrower CIs were reported by previous studies compared to VGC Analyzer whereas in one study with paired data the reported CIs were similar or even broader using ROC software. Significance testing based on the CI and the p-value from VGC Analyzer in most cases gave the same result, but sometimes led to different results in statistically weak studies when the random reader approach was used. In these cases a p-value < 0.05 was obtained although the CI of the AUCVGC included 0.5. Conclusions: The results of the present work indicate that studies using single reader adapted ROC software for analysing VGC data may, especially in non-paired data studies, underestimate the uncertainty of the obtained AUCVGC, leading to an increased risk of Type I errors. On the other hand, incorrect use of ROC software for analysis of paired data may overestimate the uncertainty of the obtained AUCVGC, leading to increased risk of Type II errors.

Page Manager: Webmaster|Last update: 9/11/2012
Share:

The University of Gothenburg uses cookies to provide you with the best possible user experience. By continuing on this website, you approve of our use of cookies.  What are cookies?