Maximum likelihood analysis of free-response receiver operating characteristic (FROC) data

Area of Science:

• Medical Imaging Analysis
• Biostatistics
• Radiology

Background:

• Receiver operating characteristic (ROC) methodology is standard for evaluating medical imaging but has limitations in abnormality detection tasks requiring localization.
• Free-response receiver operating characteristic (FROC) methodology provides a more suitable framework for observer performance studies in localization tasks.
• Lack of robust statistical analysis tools for FROC data has hindered its widespread adoption compared to ROC analysis.

Purpose of the Study:

• To develop and present a statistical model for analyzing Free-response Receiver Operating Characteristic (FROC) data.
• To introduce a computational program (FROCFIT) for parameter estimation and uncertainty quantification in FROC studies.
• To propose a performance index for quantifying observer performance in FROC experiments.

Main Methods:

• Developed a two-parameter statistical model assuming normal distributions for signal stimuli and false positive occurrences.
• Created the FROCFIT program for estimating model parameters and their uncertainties from experimental FROC data.
• Applied the methodology to existing FROC datasets to evaluate its fitting capabilities.

Main Results:

• The proposed two-parameter model effectively analyzes FROC data.
• The FROCFIT program successfully estimates parameters and uncertainties, facilitating statistical analysis.
• The developed methodology demonstrated good to excellent fits when applied to various FROC datasets.

Conclusions:

• The presented statistical model and FROCFIT program provide a viable solution for FROC data analysis.
• This methodology enhances the ability to quantify observer performance in medical imaging studies involving abnormality localization.
• The FROC methodology, supported by these statistical tools, is recommended for future observer performance evaluations in medical imaging.