Regression models for analyzing radiological visual grading studies--an empirical comparison

Area of Science:

• Medical Imaging
• Statistical Analysis
• Radiology

Background:

• Visual grading experiments are used to evaluate medical image quality.
• Observers rate image quality on an ordinal scale.
• Regression methods are available for analyzing this data.

Purpose of the Study:

• To empirically compare regression methods for analyzing ordinal image quality data.
• To evaluate the inclusion of random effects in statistical models for observer and patient variability.
• To assess different logistic regression models for medical image quality assessment.

Main Methods:

• Data from 6 observers grading 40 patients' image quality were used.
• Models tested included linear regression and various ordinal logistic regression techniques.
• Comparison focused on models with fixed effects versus mixed-effects (including random effects).

Main Results:

• Goodness-of-fit metrics (AIC, Pseudo R²) showed minor differences among fixed-effects models.
• Mixed-effects models yielded higher AIC and lower Pseudo R², potentially due to more parameters.
• Estimated dose reduction potential varied minimally across models.

Conclusions:

• Ordinal logistic regression is suggested as the most suitable method.
• This approach effectively handles ordinal data and random effects.
• It provides an appropriate framework for analyzing visual grading experiments in medical imaging.