Comparison of semiparametric receiver operating characteristic models on observer data

Area of Science:

• Medical Imaging
• Observer Performance Studies
• Statistical Modeling

Background:

• Medical imaging device evaluation relies on observer signal detection tasks.
• Receiver Operating Characteristic (ROC) models are commonly used for ordinal regression of this data.

Purpose of the Study:

• To assess the fit of various two-sample ROC models to medical imaging observer data.
• To determine the necessity of single-parameter versus multi-parameter models for describing observer performance.

Main Methods:

• Applied several two-sample ROC models to randomly selected medical imaging datasets.
• Evaluated model fit using Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), and cross-validation.

Main Results:

• A single-parameter model was found sufficient for many observer datasets.
• The power-law model, a single-parameter option, frequently provided a good fit.
• Evidence for multi-parameter models was limited in the analyzed datasets.

Conclusions:

• Single-parameter models, especially the power-law model, are often adequate for analyzing medical imaging observer data.
• The power-law model exhibits characteristics similar to the bi-normal model, including an asymmetric ROC curve.