Realistic High-Resolution Body Computed Tomography Image Synthesis by Using Progressive Growing Generative...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Generative adversarial networks (GANs) offer solutions for supervised deep learning challenges.
• Generating highly realistic synthetic images is crucial for GAN-based approaches.

Purpose of the Study:

• To validate the unsupervised synthesis of realistic body computed tomography (CT) images using a progressive growing GAN (PGGAN).
• To assess the PGGAN's ability to learn normal data distribution for image synthesis.

Main Methods:

• A PGGAN was trained on 11,755 body CT scans.
• Ten radiologists evaluated 300 images (150 real, 150 synthetic) for authenticity.
• Radiologists were categorized into three groups based on experience levels.

Main Results:

• The overall accuracy of radiologists in distinguishing real from synthetic images was 59.4%, significantly above random guessing.
• No significant difference was found in specificity for identifying synthetic images compared to random guessing (51.9% vs 50.0%).
• Accuracy did not significantly differ across radiologist experience groups, and interreader agreement was poor (κ=0.11).
• Discrepancies between real and synthetic images were most noted at the thoracoabdominal junction and in anatomical details.

Conclusions:

• GANs can synthesize highly realistic, high-resolution body CT images.
• Current GAN models have limitations in accurately depicting the thoracoabdominal junction and fine anatomical details.