Using a Visual Turing Test to Evaluate the Realism of Generative Adversarial Network (GAN)-Based Synthesized Myocardial Perfusion Images.

Abstract

As the quality of image generation by deep learning increases, it is becoming difficult to discern its authenticity from the image alone. Currently, generative models represented by generative adversarial networks (GAN) are increasingly utilized in the research field of cardiology, and their potential risks are also being pointed out. In this context, we assessed whether expert cardiologists can detect synthesized myocardial perfusion images (MPI) generated by GAN as fake. A total of 1448 polar maps collected from consecutive patients who underwent MPI for known or suspected coronary artery disease from January 2020 to December 2021 were used for the analysis. A deep convolutional GAN was trained on the polar maps to synthesize realistic MPI. The realism of the generated images in terms of human perception was evaluated by the visual Turing test (VTT) on our original website. The average correct answer rate of the VTT was only 61.1% with a standard deviation of 21.5, but this improved to 80.0±15.8 (%) in the second trial when given the clue information. In the era of machine intelligence and virtual reality, digital literacy is becoming more necessary for healthcare professionals to identify deepfakes.