Comparison of the Clinical Informativeness of Photographs and Digital Imaging Media With Multiple-Choice Receiver Operating Characteristic Analysis

Abstract

Electronic medical imaging is important for medical informatics, computerized learning, and especially for the growing field of telemedicine. The image resolution necessary for a clinical application can be determined by use of receiver operating characteristic (ROC) experiments. Completely profiling display systems is a tedious process, requiring multiple ROC experiments. We have developed a multiple-choice ROC analysis technique to compare the relative informativeness of digital image formats for a spectrum of cutaneous lesions simultaneously. The technique makes use of logical competitor sets (LCSs) of clinical conditions to redefine multiple-choice responses into the present/absent framework required for conventional ROC curve construction. The study divided 180 slides and digital images into three LCSs: pigmented lesions, flesh-colored papules, and papulosquamous conditions. Eight dermatologists diagnosed the lesions presented in two randomized viewing sessions. Accuracy profiles, independent of individual observer sensitivities, were derived from the responses. The informativeness of color slides and digital images was statistically similar, even when the conditions were stratified by difficulty of diagnosis. Results for nine specific skin conditions represented in the three LCSs were obtained simultaneously. Digital images appear to be as informative as slides for specific dermatologic diagnoses in the three LCSs tested. The use of LCSs allows stratification of results by diagnosis with greater efficiency than multiple repeated ROC experiments. Multiple-choice ROC analysis used in conjunction with logical competitor sets is the best currently available method for comparing imaging media for use in visual disciplines such as dermatology, radiology, pathology, and others.