Evaluating AI Clinically—It’s Not Just ROC AUC!
Menée à partir de clichés mammaires issus d'IRMs dynamiques avec rehaussement de contraste réalisées sur 111 femmes (âge moyen : 52 ans), cette étude évalue la performance d'un système d'intelligence artificielle pour aider les radiologues à distinguer une lésion cancéreuse d'une lésion bénigne
George Eliot once said, “History repeats itself,” but I think a twist on it attributed to Mark Twain is usually more on the mark: “The past does not repeat itself, but it often rhymes.” Research on the development, implementation, and clinical evaluation of computer-aided detection and subsequent computer-aided diagnosis schemes actually goes back further in time than most people realize—mid-1960s for computer-aided detection and 1970s for computer-aided diagnosis (1). Machine learning, deep learning, and artificial intelligence (AI), as extensions of computer-aided detection and diagnosis, have the same roots in the 1970s, although their “sudden” appearance in medical imaging did not occur until about 10 years ago. Since then, interest in development and clinical implementation has skyrocketed. The types of images, modalities, and tasks have dramatically changed compared with computer-aided diagnosis or detection applications to include applications across all aspects of the imaging enterprise from classic image analysis to report mining, billing tools, workflow management, and a host of other areas too numerous to list.
Radiology , éditorial, 2019