An Image Turing Test on Realistic Gastroscopy Images Generated by Using the Progressive Growing of Generative...

Area of Science:

• Medical imaging
• Artificial intelligence
• Gastroenterology

Background:

• Generative adversarial networks (GAN) aid in medical data augmentation and privacy.
• Existing GANs struggle to generate high-quality endoscopic images with realistic variations.
• Progressive growing of GAN (PGGAN) shows promise for medical image synthesis.

Purpose of the Study:

• To evaluate PGGAN's capability in generating high-quality gastrointestinal images.
• To identify limitations of PGGAN in creating realistic endoscopic visuals.
• To assess endoscopists' ability to differentiate synthetic from real endoscopic images.

Main Methods:

• Trained PGGAN on 107,060 normal gastroscopy images to generate 512x512 pixel images.
• Conducted visual Turing tests with 19 endoscopists comparing 100 real and 100 synthetic images.
• Analyzed endoscopist accuracy, sensitivity, and specificity based on clinical experience.

Main Results:

• Overall accuracy, sensitivity, and specificity were 61.3%, 70.3%, and 52.4%, respectively.
• No significant accuracy difference was observed across endoscopist experience groups.
• Real images with pylorus landmarks showed higher detection sensitivity; PGGAN struggled with rugal folds and mucosal texture.

Conclusions:

• PGGAN generates highly realistic gastrointestinal images, challenging for endoscopists to distinguish.
• PGGAN's ability to replicate fine details like rugal folds and mucosal texture requires improvement.
• Further GAN development is crucial for accurate endoscopic image synthesis in clinical applications.