Realistic High-Resolution Body Computed Tomography Image Synthesis by Using Progressive Growing Generative Adversarial Network: Visual Turing Test

BACKGROUND

Generative adversarial network (GAN)-based synthetic images can be viable solutions to current supervised deep learning challenges. However, generating highly realistic images is a prerequisite for these approaches.

OBJECTIVE

The aim of this study was to investigate and validate the unsupervised synthesis of highly realistic body computed tomography (CT) images by using a progressive growing GAN (PGGAN) trained to learn the probability distribution of normal data.

METHODS

We trained the PGGAN by using 11,755 body CT scans. Ten radiologists (4 radiologists with <5 years of experience [Group I], 4 radiologists with 5-10 years of experience [Group II], and 2 radiologists with >10 years of experience [Group III]) evaluated the results in a binary approach by using an independent validation set of 300 images (150 real and 150 synthetic) to judge the authenticity of each image.

RESULTS

The mean accuracy of the 10 readers in the entire image set was higher than random guessing (1781/3000, 59.4% vs 1500/3000, 50.0%, respectively; P<.001). However, in terms of identifying synthetic images as fake, there was no significant difference in the specificity between the visual Turing test and random guessing (779/1500, 51.9% vs 750/1500, 50.0%, respectively; P=.29). The accuracy between the 3 reader groups with different experience levels was not significantly different (Group I, 696/1200, 58.0%; Group II, 726/1200, 60.5%; and Group III, 359/600, 59.8%; P=.36). Interreader agreements were poor (κ=0.11) for the entire image set. In subgroup analysis, the discrepancies between real and synthetic CT images occurred mainly in the thoracoabdominal junction and in the anatomical details.

CONCLUSIONS

The GAN can synthesize highly realistic high-resolution body CT images that are indistinguishable from real images; however, it has limitations in generating body images of the thoracoabdominal junction and lacks accuracy in the anatomical details.

Challenges and Prospects in Vision and Language Research

Language grounded image understanding tasks have often been proposed as a method for evaluating progress in artificial intelligence. Ideally, these tasks should test a plethora of capabilities that integrate computer vision, reasoning, and natural language understanding. However, the datasets and evaluation procedures used in these tasks are replete with flaws which allows the vision and language (V&L) algorithms to achieve a good performance without a robust understanding of vision and language. We argue for this position based on several recent studies in V&L literature and our own observations of dataset bias, robustness, and spurious correlations. Finally, we propose that several of these challenges can be mitigated by creation of carefully designed benchmarks.