Artificial Intelligence-based Generation of Ultrasonic Images: An analysis of the detectability of synthesized images

This study uses generative artificial intelligence (AI) to synthesize ultrasonic images and investigate the detectability of AI-generated images. While generative AI can address data scarcity by generating synthetic datasets, these AI-generated images can be exploited maliciously to replace real data. Leveraging the capabilities of denoising diffusion probabilistic models (DDPMs) for synthetic image generation and convolutional neural networks (CNNs) for detecting AI-generated images, the research highlights the dual-edged nature of generative AI. A U-Net-based DDPM was trained on a dataset of 19,810 grayscale phased array ultrasonic A-scan images from weld lines in a steel pipe, producing synthetic A-scan images with a Fréchet inception distance (FID) score of 13.7. A CNN classifier, incorporating a global average pooling (GAP) layer for enhanced interpretability, was trained on both real and synthetic images to effectively differentiate between them. The results of the GAP layer visualization revealed that the CNN classifier model uses the background of the real images and the foreground of the AI-generated images to distinguish between them. This study underscores the importance of integrating robust detection mechanisms with AI advancements to authenticate and safeguard critical nondestructive evaluation (NDE) data. In addition, this work helps us understand the differences between real and DDPM-generated images providing a pathway for synthesizing more realistic images in the future.