Multiparameter Receiver Operating Characteristic Analysis for Signal Detection and Classification

Receiver operating characteristic (ROC) analysis is a widely used evaluation tool in signal processing and communications, and medical diagnosis for performance analysis. It utilizes 2-D curves plotted by detection rate (P _D) against false alarm rate (P _F) to assess effectiveness of a detector, sensor/device for detection. However, P _D and P _F are actually dependent parameters resulting from a more crucial but implicit parameter hidden in the ROC curves, threshold ¿ , which is determined by the cost of implementing a detector or sensor/device, except only the case that when the Bayes theory is used for detection, ¿ is completely determined by the Bayes cost. This paper extends the traditional ROC analysis for single-signal detection to detection and classification of multiple signals. It also explores relationships among the three parameters, P _D, P _F, and ¿ , and further develops a new concept of multiparameter ROC analysis, which uses 3-D ROC curves plotted by three parameters, P _D, P _F, and ¿, to evaluate effectiveness of detection performance based on interrelationship among P _D, P _F, and ¿, rather then only P _D and P _F used by 2-D ROC analysis. As a result of a 3-D ROC curve, three 2-D ROC curves can be also derived: the conventional 2-D ROC curve plotted by P _D versus P _F and two new 2-D ROC curves plotted based on P _D versus ¿ and P _F versus ¿. In order to demonstrate the utility of 3-D ROC analysis, four applications are considered: hyperspectral target detection, medical diagnosis, chemical/biological agent detection, and biometric recognition.