Shannon information and ROC analysis in imaging.

Area of Science:

• Medical imaging analysis
• Information theory
• Statistical decision theory

Background:

• Shannon information (SI) and receiver operating characteristic (ROC) curves are established methods for evaluating imaging system performance in binary classification tasks.
• These tasks often involve detecting signals within noisy backgrounds, crucial for medical diagnoses.

Purpose of the Study:

• To introduce a new ROC curve, the Shannon information receiver operator curve (SIROC), derived from SI for binary classification.
• To demonstrate the mathematical equivalence between the ideal-observer ROC curve and the SIROC.
• To establish a complete mathematical link between SI analysis and ideal-observer ROC analysis for imaging systems.

Main Methods:

• Derivation of the SIROC from the SI expression for binary classification tasks.
• Mathematical formulation of an integral transform to map ideal-observer ROC curves onto SIROCs.
• Development of an integral transform relating minimum probability of error to conditional entropy.

Main Results:

• The ideal-observer ROC curve and the SIROC are shown to be equivalent descriptions of optimal observer performance.
• A complete mathematical equivalence between ideal-observer ROC analysis and SI analysis is established.
• A close relationship between the areas under the ideal-observer ROC curve and the SIROC is identified, leading to new inequalities.

Conclusions:

• The SIROC provides a unified framework for analyzing imaging system performance, linking SI and ROC methodologies.
• The established mathematical equivalences offer new insights into the relationship between information theory and statistical decision theory in imaging.
• The findings facilitate a more comprehensive understanding and evaluation of imaging system performance metrics.