Reliable and computationally efficient maximum-likelihood estimation of "proper" binormal ROC curves

Area of Science:

• Medical imaging analysis
• Statistical modeling in diagnostics

Background:

• Estimating Receiver Operating Characteristic (ROC) curves in multireader, multicase (MRMC) studies presents challenges.
• Conventional methods like Wilcoxon estimates can be biased with categorical data, and standard binormal models may yield unrealistic fits.
• The proper binormal model (PBM) offers improved fits but faced numerical instability in early implementations.

Purpose of the Study:

• To introduce a novel, numerically stable algorithm for implementing the proper binormal model (PBM).
• To address limitations of previous PBM software, ensuring reliable ROC curve fitting.
• To provide a robust tool for analyzing observer performance in diagnostic studies.

Main Methods:

• Development of a new PBM curve-fitting algorithm designed for problematic datasets.
• Extensive testing on diverse simulated and real-world datasets across multiple operating systems (Windows, Linux, macOS).
• Algorithm availability as open-source software.

Main Results:

• The new algorithm demonstrated consistent convergence and produced accurate fits across millions of tested datasets.
• Reliable estimates of AUC (Area Under the Curve) standard errors were achieved for most datasets.
• AUC estimates showed favorable comparison with Wilcoxon estimates, particularly for categorical data.

Conclusions:

• The developed PBM implementation is a reliable tool for ROC curve fitting in various scenarios.
• This algorithm provides a robust solution for analyzing diagnostic accuracy in observer studies.
• The software offers improved performance for problematic datasets and categorical data analysis.